add flux2

Signed-off-by: vladmandic <mandic00@live.com>

CHANGELOG.md

@@ -1,23 +1,29 @@
 # Change Log for SD.Next
 
-## Update for 2025-11-23
+## Update for 2025-11-26
 
 ### TBD
 
 Merge commit: `f903a36d9`
 
-### Highlights for 2025-11-23
+### Highlights for 2025-11-26
 
 New native [kanvas](https://vladmandic.github.io/sdnext-docs/Kanvas/) module for image manipulation that fully replaces img2img, inpaint and outpaint controls  
-And a first cloud model with **Google's Nano Banana** *2.5 Flash and 3.0 Pro* plus new **Photoroom PRX** model  
+New generation of **Flux.2** large image model and a first cloud model with **Google's Nano Banana** *2.5 Flash and 3.0 Pro* plus new **Photoroom PRX** model  
 
 ![Screenshot](https://github.com/user-attachments/assets/54b25586-b611-4d70-a28f-ee3360944034)
 
 [ReadMe](https://github.com/vladmandic/automatic/blob/master/README.md) | [ChangeLog](https://github.com/vladmandic/automatic/blob/master/CHANGELOG.md) | [Docs](https://vladmandic.github.io/sdnext-docs/) | [WiKi](https://github.com/vladmandic/automatic/wiki) | [Discord](https://discord.com/invite/sd-next-federal-batch-inspectors-1101998836328697867) | [Sponsor](https://github.com/sponsors/vladmandic)  
 
-### Details for 2025-11-23
+### Details for 2025-11-26
 
 - **Models**
+  - [Black Forest Labs FLUX.2 Dev](https://bfl.ai/blog/flux-2) and prequantized variation [SDNQ-SVD-Uint4](https://huggingface.co/Disty0/FLUX.2-dev-SDNQ-uint4-svd-r32)  
+    **FLUX.2-Dev** is a brand new model from BFL and uses large 32B DiT together with Mistral 24B as text encoder  
+    model is available for text, image and edit tasks and can optionally use control input as second input image  
+    this is a very large model at ~100GB, so use of prequantized model at ~32GB is strongly advised
+    using prequant version and default offloading, model runs on GPUs with ~20GB  
+    *note*: model is [gated](https://vladmandic.github.io/sdnext-docs/Gated/)  
   - **Google Gemini Nano Banana** [2.5 Flash](https://blog.google/products/gemini/gemini-nano-banana-examples/) and [3.0 Pro](https://deepmind.google/models/gemini-image/pro/) 
     first cloud-based model directly supported in SD.Next UI  
     *note*: need to set `GOOGLE_API_KEY` environment variable with your key to use this model  

TODO.md

@@ -6,13 +6,10 @@
 
 ## Kanvas
 
-- server-side mask handling vs ui mask handling
-- implement different auto-masking options  
+- implement different llama remover  
 
 ## Internal
 
-- UI: New inpaint/outpaint interface  
-  [Kanvas](https://github.com/vladmandic/kanvas)  
 - Deploy: Create executable for SD.Next  
 - Feature: Integrate natural language image search  
   [ImageDB](https://github.com/vladmandic/imagedb)  
@@ -31,6 +28,7 @@
 
 ## Features
 
+- [Flux.2 TinyVAE](https://huggingface.co/fal/FLUX.2-Tiny-AutoEncoder)
 - [IPAdapter composition](https://huggingface.co/ostris/ip-composition-adapter)  
 - [IPAdapter negative guidance](https://github.com/huggingface/diffusers/discussions/7167)  
 - [MagCache](https://github.com/lllyasviel/FramePack/pull/673/files)  

html/reference.json

@@ -144,6 +144,15 @@
     "size": 32.93,
     "date": "2025 July"
   },
+  "Black Forest Labs FLUX.2 Dev": {
+    "path": "black-forest-labs/FLUX.2-dev",
+    "preview": "black-forest-labs--FLUX.2-dev.jpg",
+    "desc": "FLUX.2 generates high-quality images while maintaining character and style consistency across multiple reference images, following structured prompts, reading and writing complex text, adhering to brand guidelines, and reliably handling lighting, layouts, and logos.",
+    "skip": true,
+    "extras": "",
+    "size": 0,
+    "date": "2025 November"
+  },
   "Tencent FLUX.1 Dev SRPO": {
     "path": "vladmandic/flux.1-dev-SRPO",
     "preview": "vladmandic--flux.1-dev-SRPO.jpg",
@@ -959,6 +968,15 @@
     "date": "2025 October",
     "extras": ""
   },
+  "FLUX.2 Dev sdnq-svd-uint4": {
+    "path": "Disty0/FLUX.2-dev-SDNQ-uint4-svd-r32",
+    "preview": "black-forest-labs--FLUX.2-dev.jpg",
+    "desc": "Quantization of black-forest-labs/FLUX.2-dev using SDNQ: sdnq-svd 4-bit uint with svd rank 32",
+    "skip": true,
+    "extras": "",
+    "size": 0,
+    "date": "2025 November"
+  },
   "Chroma1-HD sdnq-svd-uint4": {
     "path": "Disty0/Chroma1-HD-SDNQ-uint4-svd-r32",
     "preview": "Disty0--Chroma1-HD-SDNQ-uint4-svd-r32.jpg",

installer.py

@@ -619,7 +619,7 @@ def check_diffusers():
     t_start = time.time()
     if args.skip_all:
         return
-    sha = 'cd3bbe2910666880307b84729176203f5785ff7e' # diffusers commit hash
+    sha = 'c8656ed73c638e51fc2e777a5fd355d69fa5220f' # diffusers commit hash
     # if args.use_rocm or args.use_zluda or args.use_directml:
     #     sha = '043ab2520f6a19fce78e6e060a68dbc947edb9f9' # lock diffusers versions for now
     pkg = pkg_resources.working_set.by_key.get('diffusers', None)

modules/modeldata.py

@@ -32,6 +32,8 @@ def get_model_type(pipe):
         model_type = 'auraflow'
     elif 'Chroma' in name:
         model_type = 'chroma'
+    elif "Flux2" in name:
+        model_type = 'f2'
     elif "Flux" in name or "Flex1" in name or "Flex2" in name:
         model_type = 'f1'
     elif "Lumina2" in name:

modules/processing_args.py

@@ -127,6 +127,8 @@ def task_specific_kwargs(p, model):
         task_args['image'] = [Image.new('RGB', (p.width, p.height), (0, 0, 0))] # monkey-patch so qwen-image-edit pipeline does not error-out on t2i
     if ('QwenImageEditPlusPipeline' in model_cls) and (p.init_control is not None) and (len(p.init_control) > 0):
         task_args['image'] += p.init_control
+    if ('Flux2' in model_cls) and (p.init_control is not None) and (len(p.init_control) > 0):
+        task_args['image'] += p.init_control
     if ('LatentConsistencyModelPipeline' in model_cls) and (len(p.init_images) > 0):
         p.ops.append('lcm')
         init_latents = [processing_vae.vae_encode(image, model=shared.sd_model, vae_type=p.vae_type).squeeze(dim=0) for image in p.init_images]
@@ -204,7 +206,7 @@ def set_pipeline_args(p, model, prompts:list, negative_prompts:list, prompts_2:t
     if (prompt_attention != 'fixed') and ('Onnx' not in model.__class__.__name__) and ('prompt' not in p.task_args) and (
         'StableDiffusion' in model.__class__.__name__ or
         'StableCascade' in model.__class__.__name__ or
-        'Flux' in model.__class__.__name__ or
+        ('Flux' in model.__class__.__name__ and 'Flux2' not in model.__class__.__name__) or
         'Chroma' in model.__class__.__name__ or
         'HiDreamImagePipeline' in model.__class__.__name__
     ):

modules/processing_helpers.py

@@ -408,7 +408,7 @@ def calculate_base_steps(p, use_denoise_start, use_refiner_start):
     if len(getattr(p, 'timesteps', [])) > 0:
         return None
     cls = shared.sd_model.__class__.__name__
-    if 'Flex' in cls or 'Kontext' in cls or 'Edit' in cls or 'Wan' in cls:
+    if 'Flex' in cls or 'Kontext' in cls or 'Edit' in cls or 'Wan' in cls or 'Flux2':
         steps = p.steps
     elif is_modular():
         steps = p.steps

modules/sd_detect.py

@@ -92,6 +92,8 @@ def guess_by_name(fn, current_guess):
         new_guess = 'HiDream'
     elif 'chroma' in fn.lower() and 'xl' not in fn.lower():
         new_guess = 'Chroma'
+    elif 'flux.2' in fn.lower():
+        new_guess = 'FLUX2'
     elif 'flux' in fn.lower() or 'flex.1' in fn.lower():
         size = round(os.path.getsize(fn) / 1024 / 1024) if os.path.isfile(fn) else 0
         if size > 11000 and size < 16000:

modules/sd_models.py

@@ -322,6 +322,10 @@ def load_diffuser_force(model_type, checkpoint_info, diffusers_load_config, op='
             from pipelines.model_flux import load_flux
             sd_model = load_flux(checkpoint_info, diffusers_load_config)
             allow_post_quant = False
+        elif model_type in ['FLUX2']:
+            from pipelines.model_flux2 import load_flux2
+            sd_model = load_flux2(checkpoint_info, diffusers_load_config)
+            allow_post_quant = False
         elif model_type in ['FLEX']:
             from pipelines.model_flex import load_flex
             sd_model = load_flex(checkpoint_info, diffusers_load_config)

modules/sd_offload.py

@@ -14,7 +14,7 @@ from modules.timer import process as process_timer
 debug = os.environ.get('SD_MOVE_DEBUG', None) is not None
 verbose = os.environ.get('SD_MOVE_VERBOSE', None) is not None
 debug_move = log.trace if debug else lambda *args, **kwargs: None
-offload_warn = ['sc', 'sd3', 'f1', 'h1', 'hunyuandit', 'auraflow', 'omnigen', 'omnigen2', 'cogview4', 'cosmos', 'chroma', 'x-omni', 'hunyuanimage', 'hunyuanimage3']
+offload_warn = ['sc', 'sd3', 'f1', 'f2', 'h1', 'hunyuandit', 'auraflow', 'omnigen', 'omnigen2', 'cogview4', 'cosmos', 'chroma', 'x-omni', 'hunyuanimage', 'hunyuanimage3']
 offload_post = ['h1']
 offload_hook_instance = None
 balanced_offload_exclude = ['CogView4Pipeline', 'MeissonicPipeline']

modules/sd_samplers_common.py

@@ -9,7 +9,7 @@ from modules import shared, devices, processing, images, sd_vae_approx, sd_vae_t
 
 SamplerData = namedtuple('SamplerData', ['name', 'constructor', 'aliases', 'options'])
 approximation_indexes = { "Simple": 0, "Approximate": 1, "TAESD": 2, "Full VAE": 3 }
-flow_models = ['f1', 'sd3', 'lumina', 'auraflow', 'sana', 'lumina2', 'cogview4', 'h1', 'cosmos', 'chroma', 'omnigen', 'omnigen2']
+flow_models = ['f1', 'f2', 'sd3', 'lumina', 'auraflow', 'sana', 'lumina2', 'cogview4', 'h1', 'cosmos', 'chroma', 'omnigen', 'omnigen2']
 warned = False
 queue_lock = threading.Lock()
 

modules/ui_common.py

@@ -419,8 +419,8 @@ def update_token_counter(text):
     from modules import extra_networks
     prompt, _ = extra_networks.parse_prompt(text)
     if shared.sd_loaded and hasattr(shared.sd_model, 'tokenizer') and shared.sd_model.tokenizer is not None:
-        has_bos_token = shared.sd_model.tokenizer.bos_token_id is not None
-        has_eos_token = shared.sd_model.tokenizer.eos_token_id is not None
+        has_bos_token = hasattr(shared.sd_model.tokenizer, 'bos_token_id') and shared.sd_model.tokenizer.bos_token_id is not None
+        has_eos_token = hasattr(shared.sd_model.tokenizer, 'eos_token_id') and shared.sd_model.tokenizer.eos_token_id is not None
         ids = shared.sd_model.tokenizer(prompt)
         ids = getattr(ids, 'input_ids', [])
         token_count = len(ids) - int(has_bos_token) - int(has_eos_token)

pipelines/model_flux2.py

@@ -0,0 +1,39 @@
+import transformers
+import diffusers
+from modules import shared, devices, sd_models, model_quant, sd_hijack_te, sd_hijack_vae
+from pipelines import generic
+
+
+def load_flux2(checkpoint_info, diffusers_load_config=None):
+    if diffusers_load_config is None:
+        diffusers_load_config = {}
+    repo_id = sd_models.path_to_repo(checkpoint_info)
+    sd_models.hf_auth_check(checkpoint_info)
+
+    load_args, _quant_args = model_quant.get_dit_args(diffusers_load_config, allow_quant=False)
+    shared.log.debug(f'Load model: type=Flux2 repo="{repo_id}" config={diffusers_load_config} offload={shared.opts.diffusers_offload_mode} dtype={devices.dtype} args={load_args}')
+
+    transformer = generic.load_transformer(repo_id, cls_name=diffusers.Flux2Transformer2DModel, load_config=diffusers_load_config)
+    text_encoder = generic.load_text_encoder(repo_id, cls_name=transformers.Mistral3ForConditionalGeneration, load_config=diffusers_load_config)
+
+    pipe = diffusers.Flux2Pipeline.from_pretrained(
+        repo_id,
+        transformer=transformer,
+        text_encoder=text_encoder,
+        cache_dir=shared.opts.diffusers_dir,
+        **load_args,
+    )
+    pipe.task_args = {
+        'output_type': 'np',
+    }
+    diffusers.pipelines.auto_pipeline.AUTO_TEXT2IMAGE_PIPELINES_MAPPING["flux2"] = diffusers.Flux2Pipeline
+    diffusers.pipelines.auto_pipeline.AUTO_IMAGE2IMAGE_PIPELINES_MAPPING["flux2"] = diffusers.Flux2Pipeline
+    diffusers.pipelines.auto_pipeline.AUTO_INPAINT_PIPELINES_MAPPING["flux2"] = diffusers.Flux2Pipeline
+
+    del text_encoder
+    del transformer
+    sd_hijack_te.init_hijack(pipe)
+    sd_hijack_vae.init_hijack(pipe)
+
+    devices.torch_gc()
+    return pipe