Public/automatic
1
0
Fork 0
mirror of https://github.com/vladmandic/automatic synced 2026-09-20 01:31:13 +02:00
Code Issues Packages Projects Releases Wiki Activity

- all guidance values set to -1 to enable using model defaults

Browse Source 
- log default values used if not overriden by user
- rename inconsistent guidance variables:

p.cfg_scale
p.image_cfg_scale -> p.cfg_image
p.diffusers_guidance_rescale -> p.cfg_rescale
p.pag_scale -> p.cfg_true
p.pag_adaptive -> p.cfg_adaptive

Signed-off-by: Vladimir Mandic <mandic00@live.com>
This commit is contained in:
Vladimir Mandic
2026-05-16 11:06:23 +02:00
parent 632032308d
commit f67562b912
30 changed files with 199 additions and 167 deletions
Download Patch File Download Diff File Expand all files Collapse all files
CHANGELOG.md
+3
View File
@@ -5,6 +5,9 @@
- **Models**
- [CircleStone Anima 1.0](https://huggingface.co/circlestone-labs/Anima) in *Base* and *Turbo* (distilled) variants
in both original precision and SDNQ-4bit quantiztion
- **Changes**
- all **Guidance** params are now set to *-1* by default to allow using model defaults and avoid confusion with different model behaviour
log will print default values used by model if not set by user
- **Features**
- **Captioning** new feature: analyze existing images for prompt adherence
*tip*: image analysis requires larger VLM model to produce quality output
data/reference-distilled.json
+1 -1
View File
@@ -14,7 +14,7 @@
"variant": "bf16",
"desc": "Stable Cascade is a diffusion model built upon the Würstchen architecture and its main difference to other models like Stable Diffusion is that it is working at a much smaller latent space. Why is this important? The smaller the latent space, the faster you can run inference and the cheaper the training becomes. How small is the latent space? Stable Diffusion uses a compression factor of 8, resulting in a 1024x1024 image being encoded to 128x128. Stable Cascade achieves a compression factor of 42, meaning that it is possible to encode a 1024x1024 image to 24x24, while maintaining crisp reconstructions. The text-conditional model is then trained in the highly compressed latent space. Previous versions of this architecture, achieved a 16x cost reduction over Stable Diffusion 1.5",
"preview": "stabilityai--stable-cascade-lite.jpg",
"extras": "sampler: Default, cfg_scale: 4.0, image_cfg_scale: 1.0",
"extras": "sampler: Default, cfg_scale: 4.0, cfg_image: 1.0",
"size": 4.97,
"tags": "distilled",
"date": "2024 February"
data/reference.json
+2 -2
View File
@@ -44,7 +44,7 @@
"variant": "bf16",
"desc": "Stable Cascade is a diffusion model built upon the W\u00fcrstchen architecture and its main difference to other models like Stable Diffusion is that it is working at a much smaller latent space. Why is this important? The smaller the latent space, the faster you can run inference and the cheaper the training becomes. How small is the latent space? Stable Diffusion uses a compression factor of 8, resulting in a 1024x1024 image being encoded to 128x128. Stable Cascade achieves a compression factor of 42, meaning that it is possible to encode a 1024x1024 image to 24x24, while maintaining crisp reconstructions. The text-conditional model is then trained in the highly compressed latent space. Previous versions of this architecture, achieved a 16x cost reduction over Stable Diffusion 1.5",
"preview": "stabilityai--stable-cascade.jpg",
"extras": "sampler: Default, cfg_scale: 4.0, image_cfg_scale: 1.0",
"extras": "sampler: Default, cfg_scale: 4.0, cfg_image: 1.0",
"size": 11.82,
"date": "2024 February"
},
@@ -823,7 +823,7 @@
"path": "warp-ai/wuerstchen",
"desc": "W\u00fcrstchen is a diffusion model whose text-conditional model works in a highly compressed latent space of images. Why is this important? Compressing data can reduce computational costs for both training and inference by magnitudes. Training on 1024x1024 images, is way more expensive than training at 32x32. Usually, other works make use of a relatively small compression, in the range of 4x - 8x spatial compression. W\u00fcrstchen takes this to an extreme. Through its novel design, we achieve a 42x spatial compression. W\u00fcrstchen employs a two-stage compression, what we call Stage A and Stage B. Stage A is a VQGAN, and Stage B is a Diffusion Autoencoder (more details can be found in the paper). A third model, Stage C, is learned in that highly compressed latent space. This training requires fractions of the compute used for current top-performing models, allowing also cheaper and faster inference.",
"preview": "warp-ai--wuerstchen.jpg",
"extras": "sampler: Default, cfg_scale: 4.0, image_cfg_scale: 0.0",
"extras": "sampler: Default, cfg_scale: 4.0, cfg_image: 0.0",
"size": 12.16,
"date": "2023 August"
},
modules/civitai/download_civitai.py
+1 -1
View File
@@ -383,7 +383,7 @@ def civitai_meta_to_parameters(meta: dict | None) -> str:
(('Size',), 'Size'),
(('Model',), 'Model'),
(('Model hash', 'modelHash'), 'Model hash'),
(('clipSkip', 'clip_skip', 'Clip skip'), 'Clip skip'),
(('clipSkip', 'clip_skip', 'CLiP-skip'), 'CLiP-skip'),
(('denoisingStrength', 'Denoising strength'), 'Denoising strength'),
]
consumed = {'prompt', 'negativeprompt', 'negative_prompt'}
modules/control/run.py
+5 -5
View File
@@ -344,7 +344,7 @@ def control_run(state: str = '', # pylint: disable=keyword-arg-before-vararg
steps: int = 20, sampler_index: int | None = None,
seed: int = -1, subseed: int = -1, subseed_strength: float = 0, seed_resize_from_h: int = -1, seed_resize_from_w: int = -1,
guidance_name: str = 'Default', guidance_scale: float = 6.0, guidance_rescale: float = 0.0, guidance_start: float = 0.0, guidance_stop: float = 1.0,
cfg_scale: float = 6.0, clip_skip: float = 1.0, image_cfg_scale: float = 6.0, diffusers_guidance_rescale: float = 0.7, pag_scale: float = 0.0, pag_adaptive: float = 0.5, cfg_end: float = 1.0,
cfg_scale: float = 6.0, clip_skip: float = 1.0, cfg_image: float = 6.0, cfg_rescale: float = 0.7, cfg_true: float = 0.0, cfg_adaptive: float = 0.5, cfg_end: float = 1.0,
vae_type: str = 'Full', tiling: bool = False, hidiffusion: bool = False,
detailer_enabled: bool = False, detailer_prompt: str = '', detailer_negative: str = '', detailer_steps: int = 10, detailer_strength: float = 0.3, detailer_resolution: int = 1024,
hdr_mode: int = 0, hdr_brightness: float = 0, hdr_color: float = 0, hdr_sharpen: float = 0, hdr_clamp: bool = False, hdr_boundary: float = 4.0, hdr_threshold: float = 0.95,
@@ -470,10 +470,10 @@ def control_run(state: str = '', # pylint: disable=keyword-arg-before-vararg
cfg_scale = cfg_scale,
cfg_end = cfg_end,
clip_skip = clip_skip,
image_cfg_scale = image_cfg_scale,
diffusers_guidance_rescale = diffusers_guidance_rescale,
pag_scale = pag_scale,
pag_adaptive = pag_adaptive,
cfg_image = cfg_image,
cfg_rescale = cfg_rescale,
cfg_true = cfg_true,
cfg_adaptive = cfg_adaptive,
# advanced
vae_type = vae_type,
tiling = tiling,
modules/image/metadata.py
+1 -1
View File
@@ -101,7 +101,7 @@ def parse_novelai_metadata(data: dict):
try:
dct = json.loads(data["Comment"])
sampler = sd_samplers.samplers_map.get(dct["sampler"], "Euler a")
geninfo = f'{data["Description"]} Negative prompt: {dct["uc"]} Steps: {dct["steps"]}, Sampler: {sampler}, CFG scale: {dct["scale"]}, Seed: {dct["seed"]}, Clip skip: 2, ENSD: 31337'
geninfo = f'{data["Description"]} Negative prompt: {dct["uc"]} Steps: {dct["steps"]}, Sampler: {sampler}, CFG scale: {dct["scale"]}, Seed: {dct["seed"]}, CLiP-skip: 2, ENSD: 31337'
debug(f'Metadata: novelai="{geninfo}"')
return geninfo
except Exception:
modules/image/namegen.py
+1 -1
View File
@@ -58,7 +58,7 @@ class FilenameGenerator:
'seed': lambda self: (hasattr(self, "seed") and self.seed and str(self.seed)) or '',
'steps': lambda self: self.p and getattr(self.p, 'steps', 0),
'cfg': lambda self: self.p and getattr(self.p, 'cfg_scale', 0),
'pag': lambda self: self.p and getattr(self.p, 'pag_scale', 0),
'pag': lambda self: self.p and getattr(self.p, 'cfg_true', 0),
'clip_skip': lambda self: self.p and getattr(self.p, 'clip_skip', 0),
'denoising': lambda self: self.p and getattr(self.p, 'denoising_strength', 0),
'styles': lambda self: (self.p and ", ".join([style for style in self.p.styles if not style == "None"])) or "None",
modules/img2img.py
+5 -5
View File
@@ -164,7 +164,7 @@ def img2img(id_task: str, state: str, mode: int,
detailer_enabled, detailer_prompt, detailer_negative, detailer_steps, detailer_strength, detailer_resolution,
n_iter, batch_size,
guidance_name, guidance_scale, guidance_rescale, guidance_start, guidance_stop,
cfg_scale, image_cfg_scale, diffusers_guidance_rescale, pag_scale, pag_adaptive, cfg_end,
cfg_scale, cfg_image, cfg_rescale, cfg_true, cfg_adaptive, cfg_end,
refiner_start,
clip_skip,
denoising_strength,
@@ -286,10 +286,10 @@ def img2img(id_task: str, state: str, mode: int,
resize_context=resize_context,
scale_by=scale_by,
denoising_strength=denoising_strength,
image_cfg_scale=image_cfg_scale,
diffusers_guidance_rescale=diffusers_guidance_rescale,
pag_scale=pag_scale,
pag_adaptive=pag_adaptive,
cfg_image=cfg_image,
cfg_rescale=cfg_rescale,
cfg_true=cfg_true,
cfg_adaptive=cfg_adaptive,
refiner_start=refiner_start,
inpaint_full_res=inpaint_full_res != 0,
inpaint_full_res_padding=inpaint_full_res_padding,
modules/ltx/ltx_process.py
+1 -1
View File
@@ -467,7 +467,7 @@ def run_ltx(task_id,
'callback_on_step_end': diffusers_callback,
'output_type': 'pil',
}
if p.cfg_scale is not None and p.cfg_scale > 0:
if p.cfg_scale is not None and p.cfg_scale > -1:
refine_args['guidance_scale'] = p.cfg_scale
if caps.supports_frame_rate_kwarg:
refine_args['frame_rate'] = float(mp4_fps)
modules/pag/__init__.py
+7 -7
View File
@@ -14,7 +14,7 @@ def apply(p: processing.StableDiffusionProcessing): # pylint: disable=arguments-
cls = shared.sd_model.__class__ if shared.sd_loaded else None
if cls == StableDiffusionPAGPipeline or cls == StableDiffusionXLPAGPipeline:
cls = unapply()
if p.pag_scale == 0:
if p.cfg_true == 0:
return
if cls is not None and 'PAG' in cls.__name__:
pass
@@ -31,20 +31,20 @@ def apply(p: processing.StableDiffusionProcessing): # pylint: disable=arguments-
orig_pipeline = shared.sd_model
shared.sd_model = sd_models.switch_pipe(StableDiffusionXLPAGPipeline, shared.sd_model)
elif detect.is_f1(cls):
p.task_args['true_cfg_scale'] = p.pag_scale
p.task_args['true_cfg_scale'] = p.cfg_true
else:
# log.warning(f'PAG: pipeline={cls.__name__} required={StableDiffusionPipeline.__name__}')
return None
p.task_args['pag_scale'] = p.pag_scale
p.task_args['pag_adaptive_scaling'] = p.pag_adaptive
p.task_args['pag_adaptive_scale'] = p.pag_adaptive
p.task_args['cfg_true'] = p.cfg_true
p.task_args['cfg_adaptive_scaling'] = p.cfg_adaptive
p.task_args['cfg_adaptive_scale'] = p.cfg_adaptive
pag_applied_layers = shared.opts.pag_apply_layers
pag_applied_layers_index = pag_applied_layers.split() if len(pag_applied_layers) > 0 else []
pag_applied_layers_index = [p.strip() for p in pag_applied_layers_index]
p.task_args['pag_applied_layers_index'] = pag_applied_layers_index if len(pag_applied_layers_index) > 0 else ['m0'] # Available layers: d[0-5], m[0], u[0-8]
p.extra_generation_params["CFG true"] = p.pag_scale
p.extra_generation_params["CFG adaptive"] = p.pag_adaptive
p.extra_generation_params["CFG true"] = p.cfg_true
p.extra_generation_params["CFG adaptive"] = p.cfg_adaptive
# log.debug(f'{c}: args={p.task_args}')
modules/pag/pipe_sd.py
+17 -17
View File
@@ -1034,20 +1034,20 @@ class StableDiffusionPAGPipeline(
return self._interrupt
@property
def pag_scale(self):
return self._pag_scale
def cfg_true(self):
return self._cfg_true
@property
def do_perturbed_attention_guidance(self):
return self._pag_scale > 0
return self._cfg_true > 0
@property
def pag_adaptive_scaling(self):
return self._pag_adaptive_scaling
def cfg_adaptive_scaling(self):
return self._cfg_adaptive_scaling
@property
def do_pag_adaptive_scaling(self):
return self._pag_adaptive_scaling > 0
def do_cfg_adaptive_scaling(self):
return self._cfg_adaptive_scaling > 0
@property
def pag_applied_layers_index(self):
@@ -1063,8 +1063,8 @@ class StableDiffusionPAGPipeline(
num_inference_steps: int = 50,
timesteps: List[int] | None = None,
guidance_scale: float = 7.5,
pag_scale: float = 0.0,
pag_adaptive_scaling: float = 0.0,
cfg_true: float = 0.0,
cfg_adaptive_scaling: float = 0.0,
pag_applied_layers_index: List[str] = ["d4"], # ['d4', 'd5', 'm0']
negative_prompt: Optional[Union[str, List[str]]] = None,
num_images_per_prompt: Optional[int] = 1,
@@ -1202,8 +1202,8 @@ class StableDiffusionPAGPipeline(
self._cross_attention_kwargs = cross_attention_kwargs
self._interrupt = False
self._pag_scale = pag_scale
self._pag_adaptive_scaling = pag_adaptive_scaling
self._cfg_true = cfg_true
self._cfg_adaptive_scaling = cfg_adaptive_scaling
self._pag_applied_layers_index = pag_applied_layers_index
# 2. Define call parameters
@@ -1372,9 +1372,9 @@ class StableDiffusionPAGPipeline(
elif not self.do_classifier_free_guidance and self.do_perturbed_attention_guidance:
noise_pred_original, noise_pred_perturb = noise_pred.chunk(2)
signal_scale = self.pag_scale
if self.do_pag_adaptive_scaling:
signal_scale = self.pag_scale - self.pag_adaptive_scaling * (1000 - t)
signal_scale = self.cfg_true
if self.do_cfg_adaptive_scaling:
signal_scale = self.cfg_true - self.cfg_adaptive_scaling * (1000 - t)
if signal_scale < 0:
signal_scale = 0
@@ -1384,9 +1384,9 @@ class StableDiffusionPAGPipeline(
elif self.do_classifier_free_guidance and self.do_perturbed_attention_guidance:
noise_pred_uncond, noise_pred_text, noise_pred_text_perturb = noise_pred.chunk(3)
signal_scale = self.pag_scale
if self.do_pag_adaptive_scaling:
signal_scale = self.pag_scale - self.pag_adaptive_scaling * (1000 - t)
signal_scale = self.cfg_true
if self.do_cfg_adaptive_scaling:
signal_scale = self.cfg_true - self.cfg_adaptive_scaling * (1000 - t)
if signal_scale < 0:
signal_scale = 0
modules/pag/pipe_sdxl.py
+17 -17
View File
@@ -1048,20 +1048,20 @@ class StableDiffusionXLPAGPipeline(
return self._interrupt
@property
def pag_scale(self):
return self._pag_scale
def cfg_true(self):
return self._cfg_true
@property
def do_adversarial_guidance(self):
return self._pag_scale > 0
return self._cfg_true > 0
@property
def pag_adaptive_scaling(self):
return self._pag_adaptive_scaling
def cfg_adaptive_scaling(self):
return self._cfg_adaptive_scaling
@property
def do_pag_adaptive_scaling(self):
return self._pag_adaptive_scaling > 0
def do_cfg_adaptive_scaling(self):
return self._cfg_adaptive_scaling > 0
@property
def pag_drop_rate(self):
@@ -1087,8 +1087,8 @@ class StableDiffusionXLPAGPipeline(
timesteps: List[int] | None = None,
denoising_end: Optional[float] = None,
guidance_scale: float = 5.0,
pag_scale: float = 0.0,
pag_adaptive_scaling: float = 0.0,
cfg_true: float = 0.0,
cfg_adaptive_scaling: float = 0.0,
pag_drop_rate: float = 0.5,
pag_applied_layers: List[str] = ['mid'], #['down', 'mid', 'up']
pag_applied_layers_index: List[str] | None = None, #['d4', 'd5', 'm0']
@@ -1308,8 +1308,8 @@ class StableDiffusionXLPAGPipeline(
self._denoising_end = denoising_end
self._interrupt = False
self._pag_scale = pag_scale
self._pag_adaptive_scaling = pag_adaptive_scaling
self._cfg_true = cfg_true
self._cfg_adaptive_scaling = cfg_adaptive_scaling
self._pag_drop_rate = pag_drop_rate
self._pag_applied_layers = pag_applied_layers
self._pag_applied_layers_index = pag_applied_layers_index
@@ -1556,9 +1556,9 @@ class StableDiffusionXLPAGPipeline(
elif not self.do_classifier_free_guidance and self.do_adversarial_guidance:
noise_pred_original, noise_pred_perturb = noise_pred.chunk(2)
signal_scale = self.pag_scale
if self.do_pag_adaptive_scaling:
signal_scale = self.pag_scale - self.pag_adaptive_scaling * (1000-t)
signal_scale = self.cfg_true
if self.do_cfg_adaptive_scaling:
signal_scale = self.cfg_true - self.cfg_adaptive_scaling * (1000-t)
if signal_scale<0:
signal_scale = 0
@@ -1569,9 +1569,9 @@ class StableDiffusionXLPAGPipeline(
noise_pred_uncond, noise_pred_text, noise_pred_text_perturb = noise_pred.chunk(3)
signal_scale = self.pag_scale
if self.do_pag_adaptive_scaling:
signal_scale = self.pag_scale - self.pag_adaptive_scaling * (1000-t)
signal_scale = self.cfg_true
if self.do_cfg_adaptive_scaling:
signal_scale = self.cfg_true - self.cfg_adaptive_scaling * (1000-t)
if signal_scale<0:
signal_scale = 0
modules/processing.py
+2 -2
View File
@@ -52,9 +52,9 @@ class Processed:
self.height = p.height if hasattr(p, 'height') else (self.images[0].height if len(self.images) > 0 else 0)
self.sampler_name = p.sampler_name or ''
self.cfg_scale = p.cfg_scale if p.cfg_scale > 1 else None
self.cfg_scale = p.cfg_scale if p.cfg_scale > -1 else None
self.cfg_end = p.cfg_end if p.cfg_end < 1 else None
self.image_cfg_scale = p.image_cfg_scale or 0
self.cfg_image = p.cfg_image if p.cfg_image > -1 else None
self.steps = p.steps or 0
self.batch_size = max(1, p.batch_size)
self.denoising_strength = p.denoising_strength
modules/processing_args.py
+27 -10
View File
@@ -179,6 +179,18 @@ def get_params(model):
return possible
def get_defaults(model, kwargs):
remove = ['return_dict', 'output_type', 'num_images_per_prompt', 'callback', 'callback_on_step_end_tensor_inputs']
try:
signature = inspect.signature(type(model).__call__, follow_wrapped=True)
defaults = {k: v.default for k, v in signature.parameters.items() if v.default is not inspect.Parameter.empty and v.default is not None} # get all defaults
defaults = {k: v for k, v in defaults.items() if k not in kwargs} # only log defaults that are not already set by kwargs
defaults = {k: v for k, v in defaults.items() if k not in remove} # remove common args that are not useful to log
log.debug(f'Pipeline: cls={model.__class__.__name__} defaults={defaults}')
except Exception as e:
log.error(f'Pipeline defaults: {e}')
def set_pipeline_args(p, model, prompts:list, negative_prompts:list, prompts_2:list | None=None, negative_prompts_2:list | None=None, prompt_attention:str | None=None, desc:str | None='', **kwargs):
t0 = time.time()
shared.sd_model = sd_models.apply_balanced_offload(shared.sd_model)
@@ -245,17 +257,20 @@ def set_pipeline_args(p, model, prompts:list, negative_prompts:list, prompts_2:l
if hasattr(model, 'scheduler') and hasattr(model.scheduler, 'noise_sampler_seed') and hasattr(model.scheduler, 'noise_sampler'):
model.scheduler.noise_sampler = None # noise needs to be reset instead of using cached values
model.scheduler.noise_sampler_seed = p.seeds # some schedulers have internal noise generator and do not use pipeline generator
if 'seed' in possible and p.seed is not None:
if ('seed' in possible) and (p.seed is not None) and (p.seed > -1):
args['seed'] = p.seed
if 'noise_sampler_seed' in possible and p.seeds is not None:
if ('noise_sampler_seed' in possible) and (p.seeds is not None):
args['noise_sampler_seed'] = p.seeds
if 'guidance_scale' in possible and p.cfg_scale is not None and p.cfg_scale > 0:
if ('guidance_scale' in possible) and (p.cfg_scale is not None) and (p.cfg_scale > -1):
args['guidance_scale'] = p.cfg_scale
if 'img_guidance_scale' in possible and hasattr(p, 'image_cfg_scale') and p.image_cfg_scale is not None and p.image_cfg_scale > 0:
args['img_guidance_scale'] = p.image_cfg_scale
if ('img_guidance_scale' in possible) and hasattr(p, 'cfg_image') and (p.cfg_image is not None) and (p.cfg_image > -1):
args['img_guidance_scale'] = p.cfg_image
if getattr(getattr(model, 'config', None), 'is_distilled', False) and args.get('guidance_scale', 0) > 1 and not getattr(p, 'distilled_warned', False):
log.warning(f'Pipeline: cls={model.__class__.__name__} distilled=True cfg_scale={args["guidance_scale"]} ignored, forced to 1')
p.distilled_warned = True
if 'generator' in possible:
generator = get_generator(p)
args['generator'] = generator
@@ -272,15 +287,14 @@ def set_pipeline_args(p, model, prompts:list, negative_prompts:list, prompts_2:l
if 'Kandinsky' in model.__class__.__name__ or 'Cosmos2' in model.__class__.__name__ or 'OmniGen2' in model.__class__.__name__:
kwargs['output_type'] = 'np' # only set latent if model has vae
if 'StableCascade' in model.__class__.__name__:
kwargs.pop("guidance_scale") # remove
kwargs.pop("num_inference_steps") # remove
if 'prior_num_inference_steps' in possible:
args["prior_num_inference_steps"] = p.steps
args["num_inference_steps"] = p.refiner_steps
if 'prior_guidance_scale' in possible:
if 'prior_guidance_scale' in possible and (p.cfg_scale is not None) and (p.cfg_scale > -1):
args["prior_guidance_scale"] = p.cfg_scale
if 'decoder_guidance_scale' in possible:
args["decoder_guidance_scale"] = p.image_cfg_scale
if 'decoder_guidance_scale' in possible and (p.cfg_image is not None) and (p.cfg_image > -1):
args["decoder_guidance_scale"] = p.cfg_image
if 'Flex2' in model.__class__.__name__:
if len(getattr(p, 'init_images', [])) > 0:
args['inpaint_image'] = p.init_images[0] if isinstance(p.init_images, list) else p.init_images
@@ -334,13 +348,14 @@ def set_pipeline_args(p, model, prompts:list, negative_prompts:list, prompts_2:l
if type(kwargs[arg]) == float or type(kwargs[arg]) == int:
if kwargs[arg] <= -1: # skip -1 as default value
continue
if kwargs[arg] is None: # skip None values
continue
args[arg] = kwargs[arg]
# optional preprocess
if hasattr(model, 'preprocess') and callable(model.preprocess):
model.preprocess(p, args)
# handle task specific args
if sd_models.get_diffusers_task(model) == sd_models.DiffusersTaskType.MODULAR:
task_kwargs = task_modular_kwargs(p, model)
@@ -402,6 +417,8 @@ def set_pipeline_args(p, model, prompts:list, negative_prompts:list, prompts_2:l
sd_hijack_hypertile.hypertile_set(p, hr=len(getattr(p, 'init_images', [])) > 0)
get_defaults(model, args)
# debug info
clean = args.copy()
clean.pop('cross_attention_kwargs', None)
modules/processing_callbacks.py
+1 -1
View File
@@ -117,7 +117,7 @@ def diffusers_callback(pipe, step: int = 0, timestep: int = 0, kwargs: dict | No
pipe._cfg_end_applied = True # pylint: disable=protected-access
if "PAG" in shared.sd_model.__class__.__name__:
pipe._guidance_scale = 1.001 if pipe._guidance_scale > 1 else pipe._guidance_scale # pylint: disable=protected-access
pipe._pag_scale = 0.001 # pylint: disable=protected-access
pipe._cfg_true = 0.001 # pylint: disable=protected-access
else:
pipe._guidance_scale = 0.0 # pylint: disable=protected-access
for key in ["prompt_embeds", "negative_prompt_embeds", "add_text_embeds", "add_time_ids"]:
modules/processing_class.py
+8 -8
View File
@@ -48,9 +48,9 @@ class StableDiffusionProcessing:
# legacy guidance
cfg_scale: float = 6.0,
cfg_end: float = 1,
diffusers_guidance_rescale: float = 0.0,
pag_scale: float = 0.0,
pag_adaptive: float = 0.5,
cfg_rescale: float = 0.0,
cfg_true: float = 0.0,
cfg_adaptive: float = 0.5,
# styles
styles: list[str] | None = None,
# vae
@@ -151,7 +151,7 @@ class StableDiffusionProcessing:
denoising_strength: float = 0.3,
init_images: list | None = None,
init_control: list | None = None,
image_cfg_scale: float | None = None,
cfg_image: float | None = None,
initial_noise_multiplier: float | None = None, # pylint: disable=unused-argument # a1111 compatibility
# resize
scale_by: float = 1,
@@ -400,7 +400,7 @@ class StableDiffusionProcessing:
self.resize_name = resize_name
self.resize_context = resize_context
self.denoising_strength = denoising_strength
self.image_cfg_scale = image_cfg_scale
self.cfg_image = cfg_image
self.scale_by = scale_by
self.mask = mask
self.image_mask = mask # TODO processing: remove duplicate mask params
@@ -465,9 +465,9 @@ class StableDiffusionProcessing:
self.guidance_stop = guidance_stop
self.cfg_scale = cfg_scale
self.cfg_end = cfg_end
self.diffusers_guidance_rescale = diffusers_guidance_rescale
self.pag_scale = pag_scale
self.pag_adaptive = pag_adaptive
self.cfg_rescale = cfg_rescale
self.cfg_true = cfg_true
self.cfg_adaptive = cfg_adaptive
self.selected_scale_tab = selected_scale_tab
self.mask_for_overlay = mask_for_overlay
self.paste_to = paste_to
modules/processing_diffusers.py
+9 -7
View File
@@ -156,9 +156,9 @@ def process_base(p: processing.StableDiffusionProcessing):
negative_prompts_2=[p.refiner_negative] if len(p.refiner_negative) > 0 else p.negative_prompts,
num_inference_steps=calculate_base_steps(p, use_refiner_start=use_refiner_start, use_denoise_start=use_denoise_start),
eta=sched_eta,
guidance_scale=p.cfg_scale,
guidance_rescale=p.diffusers_guidance_rescale,
true_cfg_scale=p.pag_scale,
guidance_scale=p.cfg_scale if p.cfg_scale is not None and p.cfg_scale > -1 else None,
guidance_rescale=p.cfg_rescale if p.cfg_rescale is not None and p.cfg_rescale > -1 else None,
true_cfg_scale=p.cfg_true if p.cfg_true is not None and p.cfg_true > -1 else None,
denoising_start=0 if use_refiner_start else p.refiner_start if use_denoise_start else None,
denoising_end=p.refiner_start if use_refiner_start else 1 if use_denoise_start else None,
num_frames=getattr(p, 'frames', 1),
@@ -323,8 +323,9 @@ def process_hires(p: processing.StableDiffusionProcessing, output):
negative_prompts_2=len(output.images) * [p.refiner_negative] if len(p.refiner_negative) > 0 else p.negative_prompts,
num_inference_steps=calculate_hires_steps(p),
eta=sched_eta,
guidance_scale=p.image_cfg_scale if p.image_cfg_scale is not None else p.cfg_scale,
guidance_rescale=p.diffusers_guidance_rescale,
guidance_scale=p.cfg_image if p.cfg_image is not None and p.cfg_image > -1 else p.cfg_scale,
guidance_rescale=p.cfg_rescale if p.cfg_rescale is not None and p.cfg_rescale > -1 else None,
true_cfg_scale=p.cfg_true if p.cfg_true is not None and p.cfg_true > -1 else None,
output_type=output_type,
clip_skip=p.clip_skip,
image=output.images,
@@ -411,8 +412,9 @@ def process_refine(p: processing.StableDiffusionProcessing, output):
num_inference_steps=calculate_refiner_steps(p),
eta=sched_eta,
noise_level=noise_level, # StableDiffusionUpscalePipeline only
guidance_scale=p.image_cfg_scale if p.image_cfg_scale is not None else p.cfg_scale,
guidance_rescale=p.diffusers_guidance_rescale,
guidance_scale=p.cfg_image if p.cfg_image is not None and p.cfg_image > -1 else p.cfg_scale,
guidance_rescale=p.cfg_rescale if p.cfg_rescale is not None and p.cfg_rescale > -1 else None,
true_cfg_scale=p.cfg_true if p.cfg_true is not None and p.cfg_true > -1 else None,
denoising_start=p.refiner_start if p.refiner_start > 0 and p.refiner_start < 1 else None,
denoising_end=1 if p.refiner_start > 0 and p.refiner_start < 1 else None,
image=image,
modules/processing_info.py
+10 -10
View File
@@ -50,12 +50,12 @@ def create_infotext(p: StableDiffusionProcessing, all_prompts=None, all_seeds=No
"Scheduler": shared.sd_model.scheduler.__class__.__name__ if getattr(shared.sd_model, 'scheduler', None) is not None else None,
"Seed": all_seeds[index],
"Seed resize from": None if p.seed_resize_from_w <= 0 or p.seed_resize_from_h <= 0 else f"{p.seed_resize_from_w}x{p.seed_resize_from_h}",
"CFG scale": p.cfg_scale if p.cfg_scale > 1.0 else 1.0,
"CFG rescale": p.diffusers_guidance_rescale if p.diffusers_guidance_rescale > 0 else None,
"CFG scale": p.cfg_scale if p.cfg_scale > -1 else None,
"CFG rescale": p.cfg_rescale if p.cfg_rescale > -1 else None,
"CFG end": p.cfg_end if p.cfg_end < 1.0 else None,
"CFG true": p.pag_scale if p.pag_scale > 0 else None,
"CFG adaptive": p.pag_adaptive if p.pag_adaptive != 0.5 else None,
"Clip skip": p.clip_skip if p.clip_skip > 1 else None,
"CFG true": p.cfg_true if p.cfg_true > 0 else None,
"CFG adaptive": p.cfg_adaptive if p.cfg_adaptive != 0.5 else None,
"CLiP-skip": p.clip_skip if p.clip_skip > 1 else None,
"Batch": f'{p.n_iter}x{p.batch_size}' if p.n_iter > 1 or p.batch_size > 1 else None,
"Refiner prompt": p.refiner_prompt if len(p.refiner_prompt) > 0 else None,
"Refiner negative": p.refiner_negative if len(p.refiner_negative) > 0 else None,
@@ -109,20 +109,20 @@ def create_infotext(p: StableDiffusionProcessing, all_prompts=None, all_seeds=No
args["Hires force"] = p.hr_force
args["Hires steps"] = p.hr_second_pass_steps
args["Hires strength"] = p.hr_denoising_strength
args["Hires sampler"] = p.hr_sampler_name
args["Hires CFG scale"] = p.image_cfg_scale
args["Hires sampler"] = p.hr_sampler_name if p.hr_sampler_name != 'Default' else None
args["Hires CFG scale"] = p.cfg_image if p.cfg_image > -1 else None
if 'refine' in p.ops:
args["Refine"] = p.enable_hr
args["Refiner"] = None if (not shared.opts.add_model_name_to_info) or (not shared.sd_refiner) or (not shared.sd_refiner.sd_checkpoint_info.model_name) else shared.sd_refiner.sd_checkpoint_info.model_name.replace(',', '').replace(':', '')
args['Hires CFG scale'] = p.image_cfg_scale
args['Hires CFG scale'] = p.cfg_image if p.cfg_image > -1 else None
args['Refiner steps'] = p.refiner_steps
args['Refiner start'] = p.refiner_start
args["Hires steps"] = p.hr_second_pass_steps
args["Hires sampler"] = p.hr_sampler_name
args["Hires sampler"] = p.hr_sampler_name if p.hr_sampler_name != 'Default' else None
if ('img2img' in p.ops or 'inpaint' in p.ops) and ('txt2img' not in p.ops and 'hires' not in p.ops): # real img2img/inpaint
args["Init image size"] = f"{getattr(p, 'init_img_width', 0)}x{getattr(p, 'init_img_height', 0)}"
args["Init image hash"] = getattr(p, 'init_img_hash', None)
args['Image CFG scale'] = p.image_cfg_scale
args['Image CFG scale'] = p.cfg_image if p.cfg_image > -1 else None
args["Mask weight"] = getattr(p, "inpainting_mask_weight", shared.opts.inpainting_mask_weight) if p.is_using_inpainting_conditioning else None
args["Denoising strength"] = getattr(p, 'denoising_strength', None)
if args["Size"] != args["Init image size"]:
modules/sd_hijack_transformers.py
+9
View File
@@ -0,0 +1,9 @@
def hijack_transformers():
# transformers>=4.56 flattened CLIPTextModel internals; diffusers single-file loader still expects `text_model`.
return
try:
import transformers
if hasattr(transformers, 'CLIPTextModel') and not hasattr(transformers.CLIPTextModel, 'text_model'):
transformers.CLIPTextModel.text_model = property(lambda self: self)
except Exception:
pass
modules/sd_models.py
+2 -1
View File
@@ -11,7 +11,7 @@ import diffusers.loaders.single_file_utils
import torch
import huggingface_hub as hf
from modules.logger import log
from modules import timer, paths, shared, shared_items, modelloader, devices, script_callbacks, sd_vae, sd_unet, errors, sd_models_compile, sd_detect, model_quant, sd_hijack_te, sd_hijack_accelerate, sd_hijack_safetensors, sd_hijack_hfhub, attention
from modules import timer, paths, shared, shared_items, modelloader, devices, script_callbacks, sd_vae, sd_unet, errors, sd_models_compile, sd_detect, model_quant, sd_hijack_te, sd_hijack_accelerate, sd_hijack_safetensors, sd_hijack_transformers, sd_hijack_hfhub, attention
from modules.memstats import memory_stats
from modules.shared_helpers import walk_files
from modules.modeldata import model_data
@@ -76,6 +76,7 @@ def set_huggingface_options(quiet=False):
else:
sd_hijack_safetensors.restore_safetensors()
sd_hijack_hfhub.init_hijack()
sd_hijack_transformers.hijack_transformers()
def set_caption_load_options():
modules/shared_legacy.py
+1 -1
View File
@@ -31,7 +31,7 @@ legacy_options = options_section(('legacy_options', "Legacy options"), {
"diffusers_move_refiner": LegacyOption(False, "Move refiner model to CPU when not in use", gr.Checkbox, {"visible": False }),
"diffusers_extract_ema": LegacyOption(False, "Use model EMA weights when possible", gr.Checkbox, {"visible": False }),
"batch_cond_uncond": LegacyOption(True, "Do conditional and unconditional denoising in one batch", gr.Checkbox, {"visible": False}),
"CLIP_stop_at_last_layers": LegacyOption(1, "Clip skip", gr.Slider, {"minimum": 1, "maximum": 8, "step": 1, "visible": False}),
"CLIP_stop_at_last_layers": LegacyOption(1, "CLiP-skip", gr.Slider, {"minimum": 1, "maximum": 8, "step": 1, "visible": False}),
"dataset_filename_join_string": LegacyOption(" ", "Filename join string", gr.Textbox, { "visible": False }),
"dataset_filename_word_regex": LegacyOption("", "Filename word regex", gr.Textbox, { "visible": False }),
"diffusers_force_zeros": LegacyOption(False, "Force zeros for prompts when empty", gr.Checkbox, {"visible": False}),
modules/txt2img.py
+5 -5
View File
@@ -17,7 +17,7 @@ def txt2img(id_task, state,
detailer_enabled, detailer_prompt, detailer_negative, detailer_steps, detailer_strength, detailer_resolution,
n_iter, batch_size,
guidance_name, guidance_scale, guidance_rescale, guidance_start, guidance_stop,
cfg_scale, image_cfg_scale, diffusers_guidance_rescale, pag_scale, pag_adaptive, cfg_end,
cfg_scale, cfg_image, cfg_rescale, cfg_true, cfg_adaptive, cfg_end,
clip_skip,
seed, subseed, subseed_strength, seed_resize_from_h, seed_resize_from_w,
height, width,
@@ -68,10 +68,10 @@ def txt2img(id_task, state,
guidance_start=guidance_start,
guidance_stop=guidance_stop,
cfg_scale=cfg_scale,
image_cfg_scale=image_cfg_scale,
diffusers_guidance_rescale=diffusers_guidance_rescale,
pag_scale=pag_scale,
pag_adaptive=pag_adaptive,
cfg_image=cfg_image,
cfg_rescale=cfg_rescale,
cfg_true=cfg_true,
cfg_adaptive=cfg_adaptive,
cfg_end=cfg_end,
clip_skip=clip_skip,
width=width,
modules/ui_common.py
+1 -1
View File
@@ -128,7 +128,7 @@ def save_files(js_data, files, html_info, index):
self.width = getattr(self, 'width', None) or getattr(self, 'Width', None) or getattr(self, 'Size-1', None) or 0
self.height = getattr(self, 'height', None) or getattr(self, 'Height', None) or getattr(self, 'Size-2', None) or 0
self.cfg_scale = getattr(self, 'cfg_scale', None) or getattr(self, 'CFG scale', None) or 0
self.clip_skip = getattr(self, 'clip_skip', None) or getattr(self, 'Clip skip', None) or 1
self.clip_skip = getattr(self, 'clip_skip', None) or getattr(self, 'CLiP-skip', None) or 1
self.denoising_strength = getattr(self, 'denoising_strength', None) or getattr(self, 'Denoising', None) or 0
self.index_of_first_image = getattr(self, 'index_of_first_image', 0)
self.subseed = getattr(self, 'subseed', None) or getattr(self, 'Subseed', None)
modules/ui_control.py
+8 -8
View File
@@ -193,7 +193,7 @@ def create_ui(_blocks: gr.Blocks=None):
mask_controls = masking.create_segment_ui()
guidance_name, guidance_scale, guidance_rescale, guidance_start, guidance_stop, cfg_scale, image_cfg_scale, diffusers_guidance_rescale, pag_scale, pag_adaptive, cfg_end = ui_guidance.create_guidance_inputs('control')
guidance_name, guidance_scale, guidance_rescale, guidance_start, guidance_stop, cfg_scale, cfg_image, cfg_rescale, cfg_true, cfg_adaptive, cfg_end = ui_guidance.create_guidance_inputs('control')
vae_type, tiling, hidiffusion, clip_skip = ui_sections.create_advanced_inputs('control')
grading_brightness, grading_contrast, grading_saturation, grading_hue, grading_gamma, grading_sharpness, grading_color_temp, grading_shadows, grading_midtones, grading_highlights, grading_clahe_clip, grading_clahe_grid, grading_shadows_tint, grading_highlights_tint, grading_split_tone_balance, grading_vignette, grading_grain, grading_lut_file, grading_lut_strength = ui_sections.create_color_inputs('control')
hdr_mode, hdr_brightness, hdr_color, hdr_sharpen, hdr_clamp, hdr_boundary, hdr_threshold, hdr_maximize, hdr_max_center, hdr_max_boundary, hdr_color_picker, hdr_tint_ratio, hdr_apply_hires = ui_sections.create_latent_inputs('control')
@@ -309,7 +309,7 @@ def create_ui(_blocks: gr.Blocks=None):
steps, sampler_index,
seed, subseed, subseed_strength, seed_resize_from_h, seed_resize_from_w,
guidance_name, guidance_scale, guidance_rescale, guidance_start, guidance_stop,
cfg_scale, clip_skip, image_cfg_scale, diffusers_guidance_rescale, pag_scale, pag_adaptive, cfg_end, vae_type, tiling, hidiffusion,
cfg_scale, clip_skip, cfg_image, cfg_rescale, cfg_true, cfg_adaptive, cfg_end, vae_type, tiling, hidiffusion,
detailer_enabled, detailer_prompt, detailer_negative, detailer_steps, detailer_strength, detailer_resolution,
hdr_mode, hdr_brightness, hdr_color, hdr_sharpen, hdr_clamp, hdr_boundary, hdr_threshold, hdr_maximize, hdr_max_center, hdr_max_boundary, hdr_color_picker, hdr_tint_ratio, hdr_apply_hires,
grading_brightness, grading_contrast, grading_saturation, grading_hue, grading_gamma, grading_sharpness, grading_color_temp,
@@ -404,10 +404,10 @@ def create_ui(_blocks: gr.Blocks=None):
# advanced
(cfg_scale, "CFG scale"),
(cfg_end, "CFG end"),
(clip_skip, "Clip skip"),
(image_cfg_scale, "Image CFG scale"),
(image_cfg_scale, "Hires CFG scale"),
(diffusers_guidance_rescale, "CFG rescale"),
(clip_skip, "CLiP-skip"),
(cfg_image, "Image CFG scale"),
(cfg_image, "Hires CFG scale"),
(cfg_rescale, "CFG rescale"),
(vae_type, "VAE type"),
(tiling, "Tiling"),
(hidiffusion, "HiDiffusion"),
@@ -438,8 +438,8 @@ def create_ui(_blocks: gr.Blocks=None):
(refiner_prompt, "Refiner prompt"),
(refiner_negative, "Refiner negative"),
# pag
(pag_scale, "CFG true"),
(pag_adaptive, "CFG adaptive"),
(cfg_true, "CFG true"),
(cfg_adaptive, "CFG adaptive"),
# hidden
(seed_resize_from_w, "Seed resize from-1"),
(seed_resize_from_h, "Seed resize from-2"),
modules/ui_guidance.py
+10 -10
View File
@@ -15,8 +15,8 @@ def create_guidance_inputs(tab):
guidance_btn = ui_components.ToolButton(value=ui_symbols.book, elem_id=f"{tab}_guider_docs")
guidance_btn.click(fn=None, _js='getGuidanceDocs', inputs=[guidance_name], outputs=[])
with gr.Row(visible=shared.opts.model_modular_enable):
guidance_scale = gr.Slider(minimum=1.0, maximum=30.0, step=0.1, label='_Guidance scale', value=4.0, elem_id=f"{tab}_guidance_scale")
guidance_rescale = gr.Slider(minimum=0.0, maximum=1.0, step=0.05, label='_Guidance rescale', value=0.0, elem_id=f"{tab}_guidance_rescale")
guidance_scale = gr.Slider(minimum=-1.0, maximum=30.0, step=0.1, label='_Guidance scale', value=-1.0, elem_id=f"{tab}_guidance_scale")
guidance_rescale = gr.Slider(minimum=-1.0, maximum=1.0, step=0.05, label='_Guidance rescale', value=-1.0, elem_id=f"{tab}_guidance_rescale")
with gr.Row(visible=shared.opts.model_modular_enable):
guidance_start = gr.Slider(minimum=0.0, maximum=1.0, step=0.05, label='_Guidance start', value=0.0, elem_id=f"{tab}_guidance_start")
guidance_stop = gr.Slider(minimum=0.0, maximum=1.0, step=0.1, label='_Guidance stop', value=1.0, elem_id=f"{tab}_guidance_stop")
@@ -55,12 +55,12 @@ def create_guidance_inputs(tab):
pag_group = gr.Accordion(open=True, label='Advanced guidance params', elem_classes=["small-accordion"], visible=False)
with pag_group:
guidance_pag_scale = gr.Slider(minimum=0.0, maximum=30.0, step=0.05, label='PAG scale', value=2.8)
guidance_cfg_true = gr.Slider(minimum=0.0, maximum=30.0, step=0.05, label='PAG scale', value=2.8)
guidance_pag_start = gr.Slider(minimum=0.0, maximum=1.0, step=0.01, label='PAG start', value=0.01)
guidance_pag_stop = gr.Slider(minimum=0.0, maximum=1.0, step=0.01, label='PAG stop', value=0.2)
guidance_pag_layers = gr.Textbox(label='PAG layers', value='7, 8, 9', placeholder='Comma-separated layer indices, e.g. 7,8,9')
guidance_pag_config = gr.Dropdown(choices=[None, 'config1', 'config2'], value=None, label='PAG config')
guidance_pag_args = [guidance_pag_scale, guidance_pag_start, guidance_pag_stop, guidance_pag_layers, guidance_pag_config]
guidance_pag_args = [guidance_cfg_true, guidance_pag_start, guidance_pag_stop, guidance_pag_layers, guidance_pag_config]
apg_group = gr.Accordion(open=True, label='Advanced guidance params', elem_classes=["small-accordion"], visible=False)
with apg_group:
@@ -114,15 +114,15 @@ def create_guidance_inputs(tab):
gr.HTML(value='<br><h2>Fallback guidance</h2>', visible=shared.opts.model_modular_enable, elem_id=f"{tab}_guidance_note")
with gr.Row(elem_id=f"{tab}_cfg_row", elem_classes=['flexbox']):
cfg_scale = gr.Slider(minimum=1.0, maximum=30.0, step=0.1, label='Guidance scale', value=6.0, elem_id=f"{tab}_cfg_scale")
cfg_scale = gr.Slider(minimum=-1.0, maximum=30.0, step=0.1, label='Guidance scale', value=-1.0, elem_id=f"{tab}_cfg_scale")
cfg_end = gr.Slider(minimum=0.0, maximum=1.0, step=0.1, label='Guidance end', value=1.0, elem_id=f"{tab}_cfg_end")
with gr.Row():
diffusers_guidance_rescale = gr.Slider(minimum=0.0, maximum=1.0, step=0.05, label='Guidance rescale', value=0.0, elem_id=f"{tab}_image_cfg_rescale")
image_cfg_scale = gr.Slider(minimum=0.0, maximum=30.0, step=0.1, label='Refine guidance', value=6.0, elem_id=f"{tab}_image_cfg_scale")
cfg_rescale = gr.Slider(minimum=-1.0, maximum=1.0, step=0.05, label='Guidance rescale', value=-1.0, elem_id=f"{tab}_image_cfg_rescale")
cfg_image = gr.Slider(minimum=-1.0, maximum=30.0, step=0.1, label='Refine guidance', value=-1.0, elem_id=f"{tab}_cfg_image")
with gr.Row():
diffusers_pag_scale = gr.Slider(minimum=0.0, maximum=30.0, step=0.05, label='Attention guidance', value=0.0, elem_id=f"{tab}_pag_scale")
diffusers_pag_adaptive = gr.Slider(minimum=0.0, maximum=1.0, step=0.05, label='Adaptive scaling', value=0.5, elem_id=f"{tab}_pag_adaptive")
cfg_true = gr.Slider(minimum=-1.0, maximum=30.0, step=0.05, label='Attention guidance', value=-1.0, elem_id=f"{tab}_cfg_true")
cfg_adaptive = gr.Slider(minimum=0.0, maximum=1.0, step=0.05, label='Adaptive scaling', value=0.5, elem_id=f"{tab}_cfg_adaptive")
_modular_args = guidance_args + lsc_args + guidance_auto_args + guidance_zero_args + guidance_pag_args + guidance_apg_args + guidance_slg_args + guidance_seg_args + guidance_fdg_args
standard_args = [cfg_scale, image_cfg_scale, diffusers_guidance_rescale, diffusers_pag_scale, diffusers_pag_adaptive, cfg_end]
standard_args = [cfg_scale, cfg_image, cfg_rescale, cfg_true, cfg_adaptive, cfg_end]
return guidance_args + standard_args
modules/ui_img2img.py
+8 -8
View File
@@ -137,7 +137,7 @@ def create_ui():
denoising_strength = gr.Slider(minimum=0.00, maximum=0.99, step=0.01, label='Denoising strength', value=0.30, elem_id="img2img_denoising_strength")
refiner_start = gr.Slider(minimum=0.0, maximum=1.0, step=0.05, label='Denoise start', value=0.0, elem_id="img2img_refiner_start")
guidance_name, guidance_scale, guidance_rescale, guidance_start, guidance_stop, cfg_scale, image_cfg_scale, diffusers_guidance_rescale, pag_scale, pag_adaptive, cfg_end = ui_guidance.create_guidance_inputs('img2img')
guidance_name, guidance_scale, guidance_rescale, guidance_start, guidance_stop, cfg_scale, cfg_image, cfg_rescale, cfg_true, cfg_adaptive, cfg_end = ui_guidance.create_guidance_inputs('img2img')
vae_type, tiling, hidiffusion, clip_skip = ui_sections.create_advanced_inputs('img2img')
grading_brightness, grading_contrast, grading_saturation, grading_hue, grading_gamma, grading_sharpness, grading_color_temp, grading_shadows, grading_midtones, grading_highlights, grading_clahe_clip, grading_clahe_grid, grading_shadows_tint, grading_highlights_tint, grading_split_tone_balance, grading_vignette, grading_grain, grading_lut_file, grading_lut_strength = ui_sections.create_color_inputs('img2img')
hdr_mode, hdr_brightness, hdr_color, hdr_sharpen, hdr_clamp, hdr_boundary, hdr_threshold, hdr_maximize, hdr_max_center, hdr_max_boundary, hdr_color_picker, hdr_tint_ratio, hdr_apply_hires = ui_sections.create_latent_inputs('img2img')
@@ -184,7 +184,7 @@ def create_ui():
detailer_enabled, detailer_prompt, detailer_negative, detailer_steps, detailer_strength, detailer_resolution,
batch_count, batch_size,
guidance_name, guidance_scale, guidance_rescale, guidance_start, guidance_stop,
cfg_scale, image_cfg_scale, diffusers_guidance_rescale, pag_scale, pag_adaptive, cfg_end,
cfg_scale, cfg_image, cfg_rescale, cfg_true, cfg_adaptive, cfg_end,
refiner_start,
clip_skip,
denoising_strength,
@@ -273,10 +273,10 @@ def create_ui():
# advanced
(cfg_scale, "CFG scale"),
(cfg_end, "CFG end"),
(image_cfg_scale, "Image CFG scale"),
(image_cfg_scale, "Hires CFG scale"),
(clip_skip, "Clip skip"),
(diffusers_guidance_rescale, "CFG rescale"),
(cfg_image, "Image CFG scale"),
(cfg_image, "Hires CFG scale"),
(clip_skip, "CLiP-skip"),
(cfg_rescale, "CFG rescale"),
(vae_type, "VAE type"),
(tiling, "Tiling"),
(hidiffusion, "HiDiffusion"),
@@ -308,8 +308,8 @@ def create_ui():
(refiner_prompt, "refiner prompt"),
(refiner_negative, "Refiner negative"),
# pag
(pag_scale, "CFG true"),
(pag_adaptive, "CFG adaptive"),
(cfg_true, "CFG true"),
(cfg_adaptive, "CFG adaptive"),
# inpaint
(mask_blur, "Mask blur"),
(mask_alpha, "Mask alpha"),
modules/ui_txt2img.py
+8 -8
View File
@@ -33,7 +33,7 @@ def create_ui():
with gr.Accordion(open=False, label="Samplers", elem_classes=["small-accordion"], elem_id="txt2img_sampler_group"):
ui_sections.create_sampler_options('txt2img')
seed, reuse_seed, subseed, reuse_subseed, subseed_strength, seed_resize_from_h, seed_resize_from_w = ui_sections.create_seed_inputs('txt2img')
guidance_name, guidance_scale, guidance_rescale, guidance_start, guidance_stop, cfg_scale, image_cfg_scale, diffusers_guidance_rescale, pag_scale, pag_adaptive, cfg_end = ui_guidance.create_guidance_inputs('txt2img')
guidance_name, guidance_scale, guidance_rescale, guidance_start, guidance_stop, cfg_scale, cfg_image, cfg_rescale, cfg_true, cfg_adaptive, cfg_end = ui_guidance.create_guidance_inputs('txt2img')
vae_type, tiling, hidiffusion, clip_skip = ui_sections.create_advanced_inputs('txt2img')
grading_brightness, grading_contrast, grading_saturation, grading_hue, grading_gamma, grading_sharpness, grading_color_temp, grading_shadows, grading_midtones, grading_highlights, grading_clahe_clip, grading_clahe_grid, grading_shadows_tint, grading_highlights_tint, grading_split_tone_balance, grading_vignette, grading_grain, grading_lut_file, grading_lut_strength = ui_sections.create_color_inputs('txt2img')
hdr_mode, hdr_brightness, hdr_color, hdr_sharpen, hdr_clamp, hdr_boundary, hdr_threshold, hdr_maximize, hdr_max_center, hdr_max_boundary, hdr_color_picker, hdr_tint_ratio, hdr_apply_hires = ui_sections.create_latent_inputs('txt2img')
@@ -59,7 +59,7 @@ def create_ui():
detailer_enabled, detailer_prompt, detailer_negative, detailer_steps, detailer_strength, detailer_resolution,
batch_count, batch_size,
guidance_name, guidance_scale, guidance_rescale, guidance_start, guidance_stop,
cfg_scale, image_cfg_scale, diffusers_guidance_rescale, pag_scale, pag_adaptive, cfg_end,
cfg_scale, cfg_image, cfg_rescale, cfg_true, cfg_adaptive, cfg_end,
clip_skip,
seed, subseed, subseed_strength, seed_resize_from_h, seed_resize_from_w,
height, width,
@@ -121,10 +121,10 @@ def create_ui():
# advanced
(cfg_scale, "CFG scale"),
(cfg_end, "CFG end"),
(clip_skip, "Clip skip"),
(image_cfg_scale, "Image CFG scale"),
(image_cfg_scale, "Hires CFG scale"),
(diffusers_guidance_rescale, "CFG rescale"),
(clip_skip, "CLiP-skip"),
(cfg_image, "Image CFG scale"),
(cfg_image, "Hires CFG scale"),
(cfg_rescale, "CFG rescale"),
(vae_type, "VAE type"),
(tiling, "Tiling"),
(hidiffusion, "HiDiffusion"),
@@ -155,8 +155,8 @@ def create_ui():
(refiner_prompt, "refiner prompt"),
(refiner_negative, "Refiner negative"),
# pag
(pag_scale, "CFG true"),
(pag_adaptive, "CFG adaptive"),
(cfg_true, "CFG true"),
(cfg_adaptive, "CFG adaptive"),
# hidden
(seed_resize_from_w, "Seed resize from-1"),
(seed_resize_from_h, "Seed resize from-2"),
modules/video_models/video_run.py
+1 -1
View File
@@ -46,7 +46,7 @@ def generate(*args, **kwargs):
denoising_strength=float(init_strength),
init_image=init_image,
cfg_scale=float(guidance_scale),
pag_scale=float(guidance_true),
cfg_true=float(guidance_true),
vae_type=vae_type,
vae_tile_frames=int(vae_tile_frames),
override_settings=override_settings,
scripts/pixelsmith/pixelsmith_pipeline.py
+23 -23
View File
@@ -1047,20 +1047,20 @@ class PixelSmithXLPipeline(
#+#
@property
def pag_scale(self):
return self._pag_scale
def cfg_true(self):
return self._cfg_true
@property
def do_adversarial_guidance(self):
return self._pag_scale > 0
return self._cfg_true > 0
@property
def pag_adaptive_scaling(self):
return self._pag_adaptive_scaling
def cfg_adaptive_scaling(self):
return self._cfg_adaptive_scaling
@property
def do_pag_adaptive_scaling(self):
return self._pag_adaptive_scaling > 0
def do_cfg_adaptive_scaling(self):
return self._cfg_adaptive_scaling > 0
@property
def pag_drop_rate(self):
@@ -1097,8 +1097,8 @@ class PixelSmithXLPipeline(
denoising_end: Optional[float] = None,
guidance_scale: float = 5.0,
#+#
pag_scale: float = 0.0, # longer inference time if used (https://ku-cvlab.github.io/Perturbed-Attention-Guidance/)
pag_adaptive_scaling: float = 0.0,
cfg_true: float = 0.0, # longer inference time if used (https://ku-cvlab.github.io/Perturbed-Attention-Guidance/)
cfg_adaptive_scaling: float = 0.0,
pag_drop_rate: float = 0.5,
pag_applied_layers: List[str] = ['mid'], #['down', 'mid', 'up']
pag_applied_layers_index: List[str] | None = None, #['d4', 'd5', 'm0']
@@ -1325,8 +1325,8 @@ class PixelSmithXLPipeline(
self._interrupt = False
#+#
self._pag_scale = pag_scale
self._pag_adaptive_scaling = pag_adaptive_scaling
self._cfg_true = cfg_true
self._cfg_adaptive_scaling = cfg_adaptive_scaling
self._pag_drop_rate = pag_drop_rate
self._pag_applied_layers = pag_applied_layers
self._pag_applied_layers_index = pag_applied_layers_index
@@ -1571,18 +1571,18 @@ class PixelSmithXLPipeline(
# pag
elif not self.do_classifier_free_guidance and self.do_adversarial_guidance:
noise_pred_original, noise_pred_perturb = noise_pred.chunk(2)
signal_scale = self.pag_scale
if self.do_pag_adaptive_scaling:
signal_scale = self.pag_scale - self.pag_adaptive_scaling * (1000-t)
signal_scale = self.cfg_true
if self.do_cfg_adaptive_scaling:
signal_scale = self.cfg_true - self.cfg_adaptive_scaling * (1000-t)
if signal_scale<0:
signal_scale = 0
noise_pred = noise_pred_original + signal_scale * (noise_pred_original - noise_pred_perturb)
# both
elif self.do_classifier_free_guidance and self.do_adversarial_guidance:
noise_pred_uncond, noise_pred_text, noise_pred_text_perturb = noise_pred.chunk(3)
signal_scale = self.pag_scale
if self.do_pag_adaptive_scaling:
signal_scale = self.pag_scale - self.pag_adaptive_scaling * (1000-t)
signal_scale = self.cfg_true
if self.do_cfg_adaptive_scaling:
signal_scale = self.cfg_true - self.cfg_adaptive_scaling * (1000-t)
if signal_scale<0:
signal_scale = 0
noise_pred = noise_pred_text + (self.guidance_scale-1.0) * (noise_pred_text - noise_pred_uncond) + signal_scale * (noise_pred_text - noise_pred_text_perturb)
@@ -1776,18 +1776,18 @@ class PixelSmithXLPipeline(
# pag
elif not self.do_classifier_free_guidance and self.do_adversarial_guidance:
noise_pred_original, noise_pred_perturb = noise_pred.chunk(2)
signal_scale = self.pag_scale
if self.do_pag_adaptive_scaling:
signal_scale = self.pag_scale - self.pag_adaptive_scaling * (1000-sub_time.max().item())
signal_scale = self.cfg_true
if self.do_cfg_adaptive_scaling:
signal_scale = self.cfg_true - self.cfg_adaptive_scaling * (1000-sub_time.max().item())
if signal_scale<0:
signal_scale = 0
noise_pred = noise_pred_original + signal_scale * (noise_pred_original - noise_pred_perturb)
# both
elif self.do_classifier_free_guidance and self.do_adversarial_guidance:
noise_pred_uncond, noise_pred_text, noise_pred_text_perturb = noise_pred.chunk(3)
signal_scale = self.pag_scale
if self.do_pag_adaptive_scaling:
signal_scale = self.pag_scale - self.pag_adaptive_scaling * (1000-sub_time.max().item())
signal_scale = self.cfg_true
if self.do_cfg_adaptive_scaling:
signal_scale = self.cfg_true - self.cfg_adaptive_scaling * (1000-sub_time.max().item())
if signal_scale<0:
signal_scale = 0
noise_pred = noise_pred_text + (self.guidance_scale-1.0) * (noise_pred_text - noise_pred_uncond) + signal_scale * (noise_pred_text - noise_pred_text_perturb)
scripts/xyz/xyz_grid_classes.py
+5 -5
View File
@@ -212,7 +212,7 @@ axis_options = [
AxisOption("[Param] Steps", int, apply_field("steps")),
AxisOption("[Param] Variation seed", int, apply_field("subseed")),
AxisOption("[Param] Variation strength", float, apply_field("subseed_strength")),
AxisOption("[Param] Clip skip", float, apply_clip_skip),
AxisOption("[Param] CLiP-skip", float, apply_clip_skip),
AxisOption("[Param] Denoising strength", float, apply_field("denoising_strength")),
AxisOptionImg2Img("[Param] Mask weight", float, apply_field("inpainting_mask_weight")),
AxisOption("[Process] Model args", str, apply_task_args),
@@ -235,8 +235,8 @@ axis_options = [
AxisOption("[Sampler] ETA multiplier", float, apply_setting("scheduler_eta")),
AxisOption("[Guidance] Scale", float, apply_field("cfg_scale")),
AxisOption("[Guidance] End", float, apply_field("cfg_end")),
AxisOption("[Guidance] Image scale", float, apply_field("image_cfg_scale")),
AxisOption("[Guidance] Rescale", float, apply_field("diffusers_guidance_rescale")),
AxisOption("[Guidance] Image scale", float, apply_field("cfg_image")),
AxisOption("[Guidance] Rescale", float, apply_field("cfg_rescale")),
AxisOption("[Guidance] Modular name", str, apply_guidance, choices=lambda: ['Default', 'CFG', 'Auto', 'Zero', 'PAG', 'APG', 'SLG', 'SEG', 'TCFG', 'FDG']),
AxisOption("[Refine] Upscaler", str, apply_field("hr_upscaler"), cost=0.3, choices=lambda: [x.name for x in shared.sd_upscalers]),
AxisOption("[Refine] Sampler", str, apply_hr_sampler_name, fmt=format_value_add_label, confirm=confirm_samplers, choices=lambda: [x.name for x in sd_samplers.samplers]),
@@ -279,8 +279,8 @@ axis_options = [
AxisOption("[HiDiffusion] T1", float, apply_override('hidiffusion_t1')),
AxisOption("[HiDiffusion] T2", float, apply_override('hidiffusion_t2')),
AxisOption("[HiDiffusion] Agression step", float, apply_field('hidiffusion_steps')),
AxisOption("[PAG] Attention scale", float, apply_field('pag_scale')),
AxisOption("[PAG] Adaptive scaling", float, apply_field('pag_adaptive')),
AxisOption("[PAG] Attention scale", float, apply_field('cfg_true')),
AxisOption("[PAG] Adaptive scaling", float, apply_field('cfg_adaptive')),
AxisOption("[PAG] Applied layers", str, apply_setting('pag_apply_layers')),
AxisOption("[IY] Scale", float, apply_task_arg('infusenet_conditioning_scale')),
AxisOption("[IY] Start", float, apply_task_arg('infusenet_guidance_start')),