add sefi-image model

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

CHANGELOG.md

@@ -3,6 +3,10 @@
 ## Update for 2026-07-31
 
 - **Models**
+  - [SeFi-Image](https://huggingface.co/SeFi-Image/SeFi-Image-5B-RL) in *Base*, *Turbo* (distilled) and *RL* (finetuned) variants  
+    SeFi is an interesting model that separates generation into semantic and texture latent streams  
+    and denoising semantic structure slightly ahead of texture details  
+    SeFi comes in 1B, 2B and 5B variants  
   - [Microsoft Mage-Flow](https://huggingface.co/mage-flow-community/Mage-Flow) in *Base* and *Turbo* (distilled) variants  
     Mage-Flow is a 4B-scale generative stack for efficient text-to-image generation and instruction-based image editing  
     *note*: Microsoft released and then unpublished the model, but we still have a mirror available for download  

data/reference-base.json

@@ -926,5 +926,37 @@
     "extras": "sampler: Default",
     "size": 16.19,
     "date": "2026 July"
+  },
+  "SeFi-Image 1B Base": {
+    "path": "vladmandic/SeFi-Image-1B-Base",
+    "preview": "vladmandic--SeFi-Image-1B-Base.jpg",
+    "desc": "SeFi-Image is a text-to-image foundation model family built with Semantic-First Diffusion. It separates generation into semantic and texture latent streams, denoising semantic structure slightly ahead of texture details.",
+    "extras": "sampler: Default",
+    "size": 0,
+    "date": "2026 July"
+  },
+  "SeFi-Image 2B Base": {
+    "path": "vladmandic/SeFi-Image-2B-Base",
+    "preview": "vladmandic--SeFi-Image-2B-Base.jpg",
+    "desc": "SeFi-Image is a text-to-image foundation model family built with Semantic-First Diffusion. It separates generation into semantic and texture latent streams, denoising semantic structure slightly ahead of texture details.",
+    "extras": "sampler: Default",
+    "size": 0,
+    "date": "2026 July"
+  },
+  "SeFi-Image 5B Base": {
+    "path": "vladmandic/SeFi-Image-5B-Base",
+    "preview": "vladmandic--SeFi-Image-5B-Base.jpg",
+    "desc": "SeFi-Image is a text-to-image foundation model family built with Semantic-First Diffusion. It separates generation into semantic and texture latent streams, denoising semantic structure slightly ahead of texture details.",
+    "extras": "sampler: Default",
+    "size": 0,
+    "date": "2026 July"
+  },
+  "SeFi-Image 5B RL": {
+    "path": "vladmandic/SeFi-Image-5B-RL",
+    "preview": "vladmandic--SeFi-Image-5B-RL.jpg",
+    "desc": "SeFi-Image is a text-to-image foundation model family built with Semantic-First Diffusion. It separates generation into semantic and texture latent streams, denoising semantic structure slightly ahead of texture details.",
+    "extras": "sampler: Default",
+    "size": 0,
+    "date": "2026 July"
   }
 }

data/reference-distilled.json

@@ -233,5 +233,29 @@
     "extras": "sampler: Default",
     "size": 16.19,
     "date": "2026 July"
+  },
+  "SeFi-Image 1B Turbo": {
+    "path": "vladmandic/SeFi-Image-1B-Turbo",
+    "preview": "vladmandic--SeFi-Image-1B-Turbo.jpg",
+    "desc": "SeFi-Image is a text-to-image foundation model family built with Semantic-First Diffusion. It separates generation into semantic and texture latent streams, denoising semantic structure slightly ahead of texture details.",
+    "extras": "sampler: Default",
+    "size": 0,
+    "date": "2026 July"
+  },
+  "SeFi-Image 2B Turbo": {
+    "path": "vladmandic/SeFi-Image-2B-Turbo",
+    "preview": "vladmandic--SeFi-Image-2B-Turbo.jpg",
+    "desc": "SeFi-Image is a text-to-image foundation model family built with Semantic-First Diffusion. It separates generation into semantic and texture latent streams, denoising semantic structure slightly ahead of texture details.",
+    "extras": "sampler: Default",
+    "size": 0,
+    "date": "2026 July"
+  },
+  "SeFi-Image 5B Turbo": {
+    "path": "vladmandic/SeFi-Image-5B-Turbo",
+    "preview": "vladmandic--SeFi-Image-5B-Turbo.jpg",
+    "desc": "SeFi-Image is a text-to-image foundation model family built with Semantic-First Diffusion. It separates generation into semantic and texture latent streams, denoising semantic structure slightly ahead of texture details.",
+    "extras": "sampler: Default",
+    "size": 0,
+    "date": "2026 July"
   }
 }

pipelines/model_anima.py

@@ -58,48 +58,6 @@ def load_anima(checkpoint_info, diffusers_load_config=None):
     if repo_id is None or repo_id.lower() == 'none':
         return None
 
-    # load-or-download custom pipeline modules from repo
-    """
-    import os
-    import sys
-    import huggingface_hub as hf
-
-    if os.path.exists(os.path.join(repo_id, 'pipeline.py')):
-        pipeline_file = os.path.join(repo_id, 'pipeline.py')
-    else:
-        try:
-            if os.path.exists(repo_id):
-                from pipelines.generic_map import transformers_map
-                custom_id = transformers_map.get('AnimaTextToImagePipeline', repo_id)
-            else:
-                custom_id = repo_id
-            pipeline_file = hf.hf_hub_download(repo_id=custom_id, filename='pipeline.py', cache_dir=shared.opts.hfcache_dir)
-        except Exception as e:
-            log.error(f'Load model: type=Anima failed to download custom modules: {e}')
-            return None
-    if os.path.exists(os.path.join(repo_id, 'llm_adapter/modeling_llm_adapter.py')):
-        adapter_file = os.path.join(repo_id, 'llm_adapter/modeling_llm_adapter.py')
-    else:
-        try:
-            if os.path.exists(repo_id):
-                from pipelines.generic_map import transformers_map
-                custom_id = transformers_map.get('AnimaTextToImagePipeline', repo_id)
-            else:
-                custom_id = repo_id
-            adapter_file = hf.hf_hub_download(repo_id=custom_id, filename='llm_adapter/modeling_llm_adapter.py', cache_dir=shared.opts.hfcache_dir)
-        except Exception as e:
-            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
-    """
-
     import sys
     from pipelines.anima import modeling_llm_adapter
     sys.modules['modeling_llm_adapter'] = modeling_llm_adapter

pipelines/model_boogu.py

@@ -22,20 +22,20 @@ def load_boogu(checkpoint_info, diffusers_load_config=None):
     from pipelines.boogu import transformer_boogu, scheduling_flow_match_euler_discrete_time_shifting
     sys.modules['transformer_boogu'] = transformer_boogu  # for loading custom code from HF repo
     sys.modules['scheduling_flow_match_euler_discrete_time_shifting'] = scheduling_flow_match_euler_discrete_time_shifting  # for loading custom code from HF repo
-    scheduler = scheduling_flow_match_euler_discrete_time_shifting.FlowMatchEulerDiscreteScheduler.from_pretrained(repo_id, subfolder='scheduler', cache_dir=shared.opts.diffusers_dir)
+    generic.set_pipeline('Boogu', BooguImagePipeline)
 
     if repo_id is None or repo_id.lower() == 'none':
         return None
 
     mllm = generic.load_text_encoder(repo_id, cls_name=transformers.Qwen3VLForConditionalGeneration, load_config=diffusers_load_config, subfolder='mllm')
     transformer = generic.load_transformer(repo_id, cls_name=BooguImageTransformer2DModel, load_config=diffusers_load_config)
+    scheduler = scheduling_flow_match_euler_discrete_time_shifting.FlowMatchEulerDiscreteScheduler.from_pretrained(repo_id, subfolder='scheduler', cache_dir=shared.opts.diffusers_dir)
 
     if 'turbo' in repo_id.lower():
         cls = BooguImageTurboPipeline
     else:
         cls = BooguImagePipeline
 
-    generic.set_pipeline('Boogu', cls)
     diffusers.pipelines.auto_pipeline.AUTO_TEXT2IMAGE_PIPELINES_MAPPING['boogu'] = cls
     diffusers.pipelines.auto_pipeline.AUTO_IMAGE2IMAGE_PIPELINES_MAPPING['boogu'] = cls
 

pipelines/model_mageflow.py

@@ -18,14 +18,15 @@ def load_mageflow(checkpoint_info, diffusers_load_config=None):
     from pipelines.mageflow import MageFlowPipeline, MageFlowTransformer2DModel
 
     log.debug(f'Load model: type=MageFlow repo="{repo_id}" config={diffusers_load_config} offload={shared.opts.diffusers_offload_mode} dtype={devices.dtype} args={diffusers_load_config}')
-    transformer = generic.load_transformer(repo_id, cls_name=MageFlowTransformer2DModel, load_config=diffusers_load_config)
-    text_encoder = generic.load_text_encoder(repo_id, cls_name=transformers.Qwen3VLForConditionalGeneration, load_config=diffusers_load_config)
-    tokenizer = transformers.Qwen2Tokenizer.from_pretrained(repo_id, subfolder="text_encoder", cache_dir=shared.opts.diffusers_dir)
+    generic.set_pipeline('MageFlow', MageFlowPipeline)
 
     if repo_id is None or repo_id.lower() == 'none':
         return None
 
-    generic.set_pipeline('MageFlow', MageFlowPipeline)
+    transformer = generic.load_transformer(repo_id, cls_name=MageFlowTransformer2DModel, load_config=diffusers_load_config)
+    text_encoder = generic.load_text_encoder(repo_id, cls_name=transformers.Qwen3VLForConditionalGeneration, load_config=diffusers_load_config)
+    tokenizer = transformers.Qwen2Tokenizer.from_pretrained(repo_id, subfolder="text_encoder", cache_dir=shared.opts.diffusers_dir)
+
     diffusers.pipelines.auto_pipeline.AUTO_TEXT2IMAGE_PIPELINES_MAPPING['mageflow'] = MageFlowPipeline
     diffusers.pipelines.auto_pipeline.AUTO_IMAGE2IMAGE_PIPELINES_MAPPING['mageflow'] = MageFlowPipeline
 

pipelines/model_prx.py

@@ -20,6 +20,7 @@ def load_prx(checkpoint_info, diffusers_load_config=None):
 
     if repo_id is None or repo_id.lower() == 'none':
         return None
+
     pipe = diffusers.PRXPipeline.from_pretrained(
         repo_id,
         transformer=transformer,

pipelines/model_sefi.py

@@ -18,6 +18,7 @@ def load_sefi(checkpoint_info, diffusers_load_config=None):
 
     transformer = generic.load_transformer(repo_id, cls_name=SeFiTransformer2DModel, load_config=diffusers_load_config)
     text_encoder = generic.load_text_encoder(repo_id, cls_name=transformers.Qwen3VLForConditionalGeneration, load_config=diffusers_load_config)
+    generic.set_pipeline('SeFi', SeFiPipeline)
     if repo_id is None or repo_id.lower() == 'none':
         return None
 
@@ -31,6 +32,10 @@ def load_sefi(checkpoint_info, diffusers_load_config=None):
 
     diffusers.pipelines.auto_pipeline.AUTO_TEXT2IMAGE_PIPELINES_MAPPING["sefi"] = SeFiPipeline
 
+    pipe.task_args = {
+        "output_type": "np",
+    }
+
     generic.load_vae_override(pipe, diffusers_load_config)
 
     del text_encoder

pipelines/sefi/__init__.py

@@ -1,3 +1,7 @@
 from .transformer_sefi import SeFiTransformer2DModel
 from .pipeline_sefi import SeFiPipeline
 from .pipeline_output import SeFiPipelineOutput
+
+import diffusers
+diffusers.SeFiTransformer2DModel = SeFiTransformer2DModel
+diffusers.SeFiPipeline = SeFiPipeline

pipelines/sefi/convert_sefi_to_diffusers.py

@@ -24,7 +24,7 @@ from huggingface_hub import snapshot_download
 from safetensors.torch import load_file
 
 from diffusers import __version__
-from diffusers.models import SeFiTransformer2DModel
+from transformer_sefi import SeFiTransformer2DModel
 
 
 SEFI_SCALE_PRESETS = {

pipelines/sefi/pipeline_sefi.py

@@ -633,7 +633,7 @@ class SeFiPipeline(DiffusionPipeline):
 
                 self._current_timestep = base_sigmas_schedule[i]
                 packed_latents = self._pack_latents(latents)
-                pred_cond = self.transformer(
+                noise_pred = self.transformer(
                     hidden_states=packed_latents,
                     timestep_sem=timesteps_sem_cur / 1000,
                     timestep_tex=timesteps_tex_cur / 1000,
@@ -643,7 +643,7 @@ class SeFiPipeline(DiffusionPipeline):
                     joint_attention_kwargs=self.attention_kwargs,
                     return_dict=False,
                 )[0]
-                pred_cond = pred_cond[:, : packed_latents.size(1)]
+                pred_cond = noise_pred[:, : packed_latents.size(1)]
                 pred_cond = self._unpack_latents_with_ids(pred_cond, latent_ids)
 
                 if self.do_classifier_free_guidance: