Merge pull request #4612 from CalamitousFelicitousness/feat/anima-pipeline

Feat/anima pipeline

data/reference-community.json

@@ -128,5 +128,12 @@
     "preview": "shuttleai--shuttle-jaguar.jpg",
     "tags": "community",
     "skip": true
+  },
+  "Anima": {
+    "path": "CalamitousFelicitousness/Anima-sdnext-diffusers",
+    "preview": "CalamitousFelicitousness--Anima-sdnext-diffusers.png",
+    "desc": "Modified Cosmos-Predict-2B that replaces the T5-11B text encoder with Qwen3-0.6B. Anima is a 2 billion parameter text-to-image model created via a collaboration between CircleStone Labs and Comfy Org. It is focused mainly on anime concepts, characters, and styles, but is also capable of generating a wide variety of other non-photorealistic content. The model is designed for making illustrations and artistic images, and will not work well at realism.",
+    "tags": "community",
+    "skip": true
   }
 }

modules/modeldata.py

@@ -58,7 +58,7 @@ def get_model_type(pipe):
         model_type = 'sana'
     elif "HiDream" in name:
         model_type = 'h1'
-    elif "Cosmos2TextToImage" in name:
+    elif "Cosmos2TextToImage" in name or "AnimaTextToImage" in name:
         model_type = 'cosmos'
     elif "FLite" in name:
         model_type = 'flite'

modules/processing_args.py

@@ -269,7 +269,7 @@ def set_pipeline_args(p, model, prompts:list, negative_prompts:list, prompts_2:t
             kwargs['output_type'] = 'np' # only set latent if model has vae
 
     # model specific
-    if 'Kandinsky' in model.__class__.__name__ or 'Cosmos2' in model.__class__.__name__ or 'OmniGen2' in model.__class__.__name__:
+    if 'Kandinsky' in model.__class__.__name__ or 'Cosmos2' in model.__class__.__name__ or 'Anima' 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

modules/sd_detect.py

@@ -103,6 +103,8 @@ def guess_by_name(fn, current_guess):
         new_guess = 'FLUX'
     elif 'flex.2' in fn.lower():
         new_guess = 'FLEX'
+    elif 'anima' in fn.lower() and 'animat' not in fn.lower():
+        new_guess = 'Anima'
     elif 'cosmos-predict2' in fn.lower():
         new_guess = 'Cosmos'
     elif 'f-lite' in fn.lower():

modules/sd_models.py

@@ -406,6 +406,10 @@ def load_diffuser_force(detected_model_type, checkpoint_info, diffusers_load_con
             from pipelines.model_cosmos import load_cosmos_t2i
             sd_model = load_cosmos_t2i(checkpoint_info, diffusers_load_config)
             allow_post_quant = False
+        elif model_type in ['Anima']:
+            from pipelines.model_anima import load_anima
+            sd_model = load_anima(checkpoint_info, diffusers_load_config)
+            allow_post_quant = False
         elif model_type in ['FLite']:
             from pipelines.model_flite import load_flite
             sd_model = load_flite(checkpoint_info, diffusers_load_config)
@@ -1248,6 +1252,8 @@ def set_diffuser_pipe(pipe, new_pipe_type):
 def add_noise_pred_to_diffusers_callback(pipe):
     if not hasattr(pipe, "_callback_tensor_inputs"):
         return pipe
+    if pipe.__class__.__name__.startswith("Anima"):
+        return pipe
     if pipe.__class__.__name__.startswith("StableCascade") and ("predicted_image_embedding" not in pipe._callback_tensor_inputs): # pylint: disable=protected-access
         pipe.prior_pipe._callback_tensor_inputs.append("predicted_image_embedding") # pylint: disable=protected-access
     elif "noise_pred" not in pipe._callback_tensor_inputs: # pylint: disable=protected-access

modules/shared_items.py

@@ -64,6 +64,7 @@ pipelines = {
     'X-Omni': getattr(diffusers, 'DiffusionPipeline', None),
     'HunyuanImage3': getattr(diffusers, 'DiffusionPipeline', None),
     'ChronoEdit': getattr(diffusers, 'DiffusionPipeline', None),
+    'Anima': getattr(diffusers, 'DiffusionPipeline', None),
 }
 
 

pipelines/model_anima.py

@@ -0,0 +1,85 @@
+import sys
+import importlib.util
+import transformers
+import diffusers
+import huggingface_hub as hf
+from modules import shared, devices, sd_models, model_quant, sd_hijack_te, sd_hijack_vae
+from pipelines import generic
+
+
+def _import_from_file(module_name, file_path):
+    spec = importlib.util.spec_from_file_location(module_name, file_path)
+    mod = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(mod)
+    return mod
+
+
+def load_anima(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=Anima repo="{repo_id}" config={diffusers_load_config} offload={shared.opts.diffusers_offload_mode} dtype={devices.dtype} args={load_args}')
+
+    # download custom pipeline modules from repo
+    try:
+        pipeline_file = hf.hf_hub_download(repo_id, filename='pipeline.py', cache_dir=shared.opts.diffusers_dir)
+        adapter_file = hf.hf_hub_download(repo_id, filename='llm_adapter/modeling_llm_adapter.py', cache_dir=shared.opts.diffusers_dir)
+    except Exception as e:
+        shared.log.error(f'Load model: type=Anima failed to download custom modules: {e}')
+        return None
+
+    # dynamically import custom classes and register in sys.modules so
+    # Diffusers' from_pretrained can resolve them via trust_remote_code
+    adapter_mod = _import_from_file('modeling_llm_adapter', adapter_file)
+    sys.modules['modeling_llm_adapter'] = adapter_mod
+    pipeline_mod = _import_from_file('pipeline', pipeline_file)
+    sys.modules['pipeline'] = pipeline_mod
+    AnimaTextToImagePipeline = pipeline_mod.AnimaTextToImagePipeline
+    AnimaLLMAdapter = adapter_mod.AnimaLLMAdapter
+
+    # load components
+    transformer = generic.load_transformer(repo_id, cls_name=diffusers.CosmosTransformer3DModel, load_config=diffusers_load_config, subfolder="transformer")
+    text_encoder = generic.load_text_encoder(repo_id, cls_name=transformers.Qwen3Model, load_config=diffusers_load_config, subfolder="text_encoder", allow_shared=False)
+
+    shared.state.begin('Load adapter')
+    try:
+        llm_adapter = AnimaLLMAdapter.from_pretrained(
+            repo_id,
+            subfolder="llm_adapter",
+            cache_dir=shared.opts.diffusers_dir,
+            torch_dtype=devices.dtype,
+        )
+    except Exception as e:
+        shared.log.error(f'Load model: type=Anima adapter: {e}')
+        return None
+    finally:
+        shared.state.end()
+
+    tokenizer = transformers.AutoTokenizer.from_pretrained(repo_id, subfolder="tokenizer", cache_dir=shared.opts.diffusers_dir)
+    t5_tokenizer = transformers.AutoTokenizer.from_pretrained(repo_id, subfolder="t5_tokenizer", cache_dir=shared.opts.diffusers_dir)
+
+    # assemble pipeline
+    pipe = AnimaTextToImagePipeline.from_pretrained(
+        repo_id,
+        transformer=transformer,
+        text_encoder=text_encoder,
+        llm_adapter=llm_adapter,
+        tokenizer=tokenizer,
+        t5_tokenizer=t5_tokenizer,
+        cache_dir=shared.opts.diffusers_dir,
+        trust_remote_code=True,
+        **load_args,
+    )
+
+    del text_encoder
+    del transformer
+    del llm_adapter
+
+    sd_hijack_te.init_hijack(pipe)
+    sd_hijack_vae.init_hijack(pipe)
+
+    devices.torch_gc()
+    return pipe