feat: add Anima (Cosmos-Predict-2B variant) pipeline support

Anima replaces the Cosmos T5-11B text encoder with Qwen3-0.6B + a
6-layer LLM adapter and uses CONST preconditioning instead of EDM.

- Add pipelines/model_anima.py loader with dynamic import of custom
  AnimaTextToImagePipeline and AnimaLLMAdapter from model repo
- Register 'Anima' pipeline in shared_items.py
- Add name-based detection in sd_detect.py
- Fix list-format _class_name handling in guess_by_diffusers()
- Wire loader in sd_models.py load_diffuser_force()
- Skip noise_pred callback injection for Anima (uses velocity instead)
- Add output_type='np' override in processing_args.py

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 'cosmos' 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 +1,83 @@
+import os
+import importlib.util
+import transformers
+import diffusers
+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}')
+
+    # resolve local path for custom pipeline modules
+    local_path = sd_models.path_to_repo(checkpoint_info, local=True)
+    pipeline_file = os.path.join(local_path, 'pipeline.py')
+    adapter_file = os.path.join(local_path, 'llm_adapter', 'modeling_llm_adapter.py')
+    if not os.path.isfile(pipeline_file):
+        shared.log.error(f'Load model: type=Anima missing pipeline.py in "{local_path}"')
+        return None
+    if not os.path.isfile(adapter_file):
+        shared.log.error(f'Load model: type=Anima missing llm_adapter/modeling_llm_adapter.py in "{local_path}"')
+        return None
+
+    # dynamically import custom classes from the model repo
+    pipeline_mod = _import_from_file('anima_pipeline', pipeline_file)
+    adapter_mod = _import_from_file('anima_llm_adapter', adapter_file)
+    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,
+        **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