support Mugen

backend/diffusion_engine/anima.py

@@ -15,7 +15,6 @@ class Anima(ForgeDiffusionEngine):
 
     def __init__(self, estimated_config, huggingface_components):
         super().__init__(estimated_config, huggingface_components)
-        self.is_inpaint = False
 
         clip = CLIP(model_dict={"qwen3_06b": huggingface_components["text_encoder"]}, tokenizer_dict={"qwen3_06b": huggingface_components["tokenizer"], "t5xxl": huggingface_components["tokenizer_2"]})
 

backend/diffusion_engine/chroma.py

@@ -15,7 +15,6 @@ class Chroma(ForgeDiffusionEngine):
 
     def __init__(self, estimated_config, huggingface_components):
         super().__init__(estimated_config, huggingface_components)
-        self.is_inpaint = False
 
         clip = CLIP(model_dict={"t5xxl": huggingface_components["text_encoder"]}, tokenizer_dict={"t5xxl": huggingface_components["tokenizer"]})
 

backend/diffusion_engine/flux.py

@@ -22,7 +22,6 @@ class Flux(ForgeDiffusionEngine):
 
     def __init__(self, estimated_config, huggingface_components):
         super().__init__(estimated_config, huggingface_components)
-        self.is_inpaint = False
 
         clip = CLIP(model_dict={"clip_l": huggingface_components["text_encoder"], "t5xxl": huggingface_components["text_encoder_2"]}, tokenizer_dict={"clip_l": huggingface_components["tokenizer"], "t5xxl": huggingface_components["tokenizer_2"]})
 

backend/diffusion_engine/flux2.py

@@ -21,7 +21,6 @@ class Flux2(ForgeDiffusionEngine):
 
     def __init__(self, estimated_config, huggingface_components):
         super().__init__(estimated_config, huggingface_components)
-        self.is_inpaint = False
 
         clip = CLIP(model_dict={"qwen3": huggingface_components["text_encoder"]}, tokenizer_dict={"qwen3": huggingface_components["tokenizer"]})
 

backend/diffusion_engine/lumina.py

@@ -15,7 +15,6 @@ class Lumina2(ForgeDiffusionEngine):
 
     def __init__(self, estimated_config, huggingface_components):
         super().__init__(estimated_config, huggingface_components)
-        self.is_inpaint = False
 
         clip = CLIP(model_dict={"gemma2": huggingface_components["text_encoder"]}, tokenizer_dict={"gemma2": huggingface_components["tokenizer"]})
 

backend/diffusion_engine/mugen.py

@@ -0,0 +1,123 @@
+import torch
+from huggingface_guess import model_list
+
+from backend import memory_management
+from backend.args import dynamic_args
+from backend.diffusion_engine.base import ForgeDiffusionEngine, ForgeObjects
+from backend.modules.k_prediction import PredictionDiscreteFlow
+from backend.nn.unet import Timestep
+from backend.patcher.clip import CLIP
+from backend.patcher.unet import UnetPatcher
+from backend.patcher.vae import VAE
+from backend.text_processing.classic_engine import ClassicTextProcessingEngine
+from modules.shared import opts
+
+
+class Mugen(ForgeDiffusionEngine):
+    matched_guesses = [model_list.Mugen]
+
+    def __init__(self, estimated_config, huggingface_components):
+        super().__init__(estimated_config, huggingface_components)
+
+        clip = CLIP(model_dict={"clip_l": huggingface_components["text_encoder"], "clip_g": huggingface_components["text_encoder_2"]}, tokenizer_dict={"clip_l": huggingface_components["tokenizer"], "clip_g": huggingface_components["tokenizer_2"]})
+
+        vae = VAE(model=huggingface_components["vae"], is_mugen=True)
+
+        k_predictor = PredictionDiscreteFlow(estimated_config)
+        unet = UnetPatcher.from_model(model=huggingface_components["unet"], diffusers_scheduler=None, k_predictor=k_predictor, config=estimated_config)
+
+        self.text_processing_engine_l = ClassicTextProcessingEngine(
+            text_encoder=clip.cond_stage_model.clip_l,
+            tokenizer=clip.tokenizer.clip_l,
+            embedding_dir=dynamic_args.embedding_dir,
+            embedding_key="clip_l",
+            embedding_expected_shape=2048,
+            text_projection=False,
+            minimal_clip_skip=2,
+            clip_skip=2,
+            return_pooled=False,
+            final_layer_norm=False,
+        )
+
+        self.text_processing_engine_g = ClassicTextProcessingEngine(
+            text_encoder=clip.cond_stage_model.clip_g,
+            tokenizer=clip.tokenizer.clip_g,
+            embedding_dir=dynamic_args.embedding_dir,
+            embedding_key="clip_g",
+            embedding_expected_shape=2048,
+            text_projection=True,
+            minimal_clip_skip=2,
+            clip_skip=2,
+            return_pooled=True,
+            final_layer_norm=False,
+        )
+
+        self.embedder = Timestep(256)
+
+        self.forge_objects = ForgeObjects(unet=unet, clip=clip, vae=vae, clipvision=None)
+        self.forge_objects_original = self.forge_objects.shallow_copy()
+        self.forge_objects_after_applying_lora = self.forge_objects.shallow_copy()
+
+        # WebUI Legacy
+        self.is_sdxl = True
+
+    def set_clip_skip(self, clip_skip):
+        self.text_processing_engine_l.clip_skip = clip_skip
+        self.text_processing_engine_g.clip_skip = clip_skip
+
+    @torch.inference_mode()
+    def get_learned_conditioning(self, prompt: list[str]):
+        memory_management.load_model_gpu(self.forge_objects.clip.patcher)
+
+        shift = getattr(prompt, "distilled_cfg_scale", 3.0)
+        self.forge_objects.unet.model.predictor.set_parameters(shift=shift)
+        memory_management.logger.debug(f"Shift: {shift}")
+
+        cond_l = self.text_processing_engine_l(prompt)
+        cond_g, clip_pooled = self.text_processing_engine_g(prompt)
+
+        width = getattr(prompt, "width", 1024) or 1024
+        height = getattr(prompt, "height", 1024) or 1024
+
+        crop_w = opts.sdxl_crop_left
+        crop_h = opts.sdxl_crop_top
+        target_width = width
+        target_height = height
+
+        out = [self.embedder(torch.Tensor([height])), self.embedder(torch.Tensor([width])), self.embedder(torch.Tensor([crop_h])), self.embedder(torch.Tensor([crop_w])), self.embedder(torch.Tensor([target_height])), self.embedder(torch.Tensor([target_width]))]
+
+        flat = torch.flatten(torch.cat(out)).unsqueeze(dim=0).repeat(clip_pooled.shape[0], 1).to(clip_pooled)
+
+        if opts.sdxl_zero_neg and getattr(prompt, "is_negative_prompt", False) and all(x == "" for x in prompt):
+            clip_pooled = torch.zeros_like(clip_pooled)
+            cond_l = torch.zeros_like(cond_l)
+            cond_g = torch.zeros_like(cond_g)
+
+        # ensure cond_l and cond_g have the same length
+        max_len = max(cond_l.shape[1], cond_g.shape[1])
+        cond_l = torch.cat([cond_l, cond_l.new_zeros(cond_l.size(0), max_len - cond_l.shape[1], cond_l.size(2))], dim=1)
+        cond_g = torch.cat([cond_g, cond_g.new_zeros(cond_g.size(0), max_len - cond_g.shape[1], cond_g.size(2))], dim=1)
+
+        cond = dict(
+            crossattn=torch.cat([cond_l, cond_g], dim=2),
+            vector=torch.cat([clip_pooled, flat], dim=1),
+        )
+
+        return cond
+
+    @torch.inference_mode()
+    def get_prompt_lengths_on_ui(self, prompt):
+        _, token_count = self.text_processing_engine_l.process_texts([prompt])
+        return token_count, self.text_processing_engine_l.get_target_prompt_token_count(token_count)
+
+    @torch.inference_mode()
+    def encode_first_stage(self, x):
+        sample = self.forge_objects.vae.encode(x.movedim(1, -1) * 0.5 + 0.5)
+        sample = self.forge_objects.vae.first_stage_model.process_in(sample)
+        return sample.to(x)
+
+    @torch.inference_mode()
+    def decode_first_stage(self, x):
+        sample = self.forge_objects.vae.first_stage_model.process_out(x)
+        sample = self.forge_objects.vae.decode(sample).movedim(-1, 1) * 2.0 - 1.0
+        return sample.to(x)

backend/diffusion_engine/qwen.py

@@ -23,7 +23,6 @@ class QwenImage(ForgeDiffusionEngine):
 
     def __init__(self, estimated_config, huggingface_components):
         super().__init__(estimated_config, huggingface_components)
-        self.is_inpaint = False
 
         clip = CLIP(model_dict={"qwen25_7b": huggingface_components["text_encoder"]}, tokenizer_dict={"qwen25_7b": huggingface_components["tokenizer"]})
 

backend/diffusion_engine/wan.py

@@ -20,7 +20,6 @@ class Wan(ForgeDiffusionEngine):
 
     def __init__(self, estimated_config, huggingface_components):
         super().__init__(estimated_config, huggingface_components)
-        self.is_inpaint = False
 
         clip = CLIP(model_dict={"umt5xxl": huggingface_components["text_encoder"]}, tokenizer_dict={"umt5xxl": huggingface_components["tokenizer"]})
 

backend/diffusion_engine/zimage.py

@@ -15,7 +15,6 @@ class ZImage(ForgeDiffusionEngine):
 
     def __init__(self, estimated_config, huggingface_components):
         super().__init__(estimated_config, huggingface_components)
-        self.is_inpaint = False
 
         clip = CLIP(model_dict={"qwen3": huggingface_components["text_encoder"]}, tokenizer_dict={"qwen3": huggingface_components["tokenizer"]})
 

backend/huggingface/CabalResearch/Mugen/model_index.json

@@ -0,0 +1,34 @@
+{
+  "_class_name": "StableDiffusionXLPipeline",
+  "_diffusers_version": "0.19.0.dev0",
+  "force_zeros_for_empty_prompt": true,
+  "add_watermarker": null,
+  "scheduler": [
+    "diffusers",
+    "EulerDiscreteScheduler"
+  ],
+  "text_encoder": [
+    "transformers",
+    "CLIPTextModel"
+  ],
+  "text_encoder_2": [
+    "transformers",
+    "CLIPTextModelWithProjection"
+  ],
+  "tokenizer": [
+    "transformers",
+    "CLIPTokenizer"
+  ],
+  "tokenizer_2": [
+    "transformers",
+    "CLIPTokenizer"
+  ],
+  "unet": [
+    "diffusers",
+    "UNet2DConditionModel"
+  ],
+  "vae": [
+    "diffusers",
+    "AutoencoderKLFlux2"
+  ]
+}

backend/huggingface/CabalResearch/Mugen/scheduler/scheduler_config.json

@@ -0,0 +1,18 @@
+{
+  "_class_name": "EulerDiscreteScheduler",
+  "_diffusers_version": "0.19.0.dev0",
+  "beta_end": 0.012,
+  "beta_schedule": "scaled_linear",
+  "beta_start": 0.00085,
+  "clip_sample": false,
+  "interpolation_type": "linear",
+  "num_train_timesteps": 1000,
+  "prediction_type": "epsilon",
+  "sample_max_value": 1.0,
+  "set_alpha_to_one": false,
+  "skip_prk_steps": true,
+  "steps_offset": 1,
+  "timestep_spacing": "leading",
+  "trained_betas": null,
+  "use_karras_sigmas": false
+}

backend/huggingface/CabalResearch/Mugen/text_encoder/config.json

@@ -0,0 +1,24 @@
+{
+  "architectures": [
+    "CLIPTextModel"
+  ],
+  "attention_dropout": 0.0,
+  "bos_token_id": 0,
+  "dropout": 0.0,
+  "eos_token_id": 2,
+  "hidden_act": "quick_gelu",
+  "hidden_size": 768,
+  "initializer_factor": 1.0,
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "layer_norm_eps": 1e-05,
+  "max_position_embeddings": 77,
+  "model_type": "clip_text_model",
+  "num_attention_heads": 12,
+  "num_hidden_layers": 12,
+  "pad_token_id": 1,
+  "projection_dim": 768,
+  "torch_dtype": "float16",
+  "transformers_version": "4.32.0.dev0",
+  "vocab_size": 49408
+}

backend/huggingface/CabalResearch/Mugen/text_encoder_2/config.json

@@ -0,0 +1,24 @@
+{
+  "architectures": [
+    "CLIPTextModelWithProjection"
+  ],
+  "attention_dropout": 0.0,
+  "bos_token_id": 0,
+  "dropout": 0.0,
+  "eos_token_id": 2,
+  "hidden_act": "gelu",
+  "hidden_size": 1280,
+  "initializer_factor": 1.0,
+  "initializer_range": 0.02,
+  "intermediate_size": 5120,
+  "layer_norm_eps": 1e-05,
+  "max_position_embeddings": 77,
+  "model_type": "clip_text_model",
+  "num_attention_heads": 20,
+  "num_hidden_layers": 32,
+  "pad_token_id": 1,
+  "projection_dim": 1280,
+  "torch_dtype": "float16",
+  "transformers_version": "4.32.0.dev0",
+  "vocab_size": 49408
+}

backend/huggingface/CabalResearch/Mugen/tokenizer/special_tokens_map.json

@@ -0,0 +1,24 @@
+{
+  "bos_token": {
+    "content": "<|startoftext|>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": "<|endoftext|>",
+  "unk_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  }
+}

backend/huggingface/CabalResearch/Mugen/tokenizer/tokenizer_config.json

@@ -0,0 +1,33 @@
+{
+  "add_prefix_space": false,
+  "bos_token": {
+    "__type": "AddedToken",
+    "content": "<|startoftext|>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "clean_up_tokenization_spaces": true,
+  "do_lower_case": true,
+  "eos_token": {
+    "__type": "AddedToken",
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "errors": "replace",
+  "model_max_length": 77,
+  "pad_token": "<|endoftext|>",
+  "tokenizer_class": "CLIPTokenizer",
+  "unk_token": {
+    "__type": "AddedToken",
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  }
+}

backend/huggingface/CabalResearch/Mugen/tokenizer_2/special_tokens_map.json

@@ -0,0 +1,24 @@
+{
+  "bos_token": {
+    "content": "<|startoftext|>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": "!",
+  "unk_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  }
+}

backend/huggingface/CabalResearch/Mugen/tokenizer_2/tokenizer_config.json

@@ -0,0 +1,33 @@
+{
+  "add_prefix_space": false,
+  "bos_token": {
+    "__type": "AddedToken",
+    "content": "<|startoftext|>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "clean_up_tokenization_spaces": true,
+  "do_lower_case": true,
+  "eos_token": {
+    "__type": "AddedToken",
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "errors": "replace",
+  "model_max_length": 77,
+  "pad_token": "!",
+  "tokenizer_class": "CLIPTokenizer",
+  "unk_token": {
+    "__type": "AddedToken",
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  }
+}

backend/huggingface/CabalResearch/Mugen/unet/config.json

@@ -0,0 +1,69 @@
+{
+  "_class_name": "UNet2DConditionModel",
+  "_diffusers_version": "0.19.0.dev0",
+  "act_fn": "silu",
+  "addition_embed_type": "text_time",
+  "addition_embed_type_num_heads": 64,
+  "addition_time_embed_dim": 256,
+  "attention_head_dim": [
+    5,
+    10,
+    20
+  ],
+  "block_out_channels": [
+    320,
+    640,
+    1280
+  ],
+  "center_input_sample": false,
+  "class_embed_type": null,
+  "class_embeddings_concat": false,
+  "conv_in_kernel": 3,
+  "conv_out_kernel": 3,
+  "cross_attention_dim": 2048,
+  "cross_attention_norm": null,
+  "down_block_types": [
+    "DownBlock2D",
+    "CrossAttnDownBlock2D",
+    "CrossAttnDownBlock2D"
+  ],
+  "downsample_padding": 1,
+  "dual_cross_attention": false,
+  "encoder_hid_dim": null,
+  "encoder_hid_dim_type": null,
+  "flip_sin_to_cos": true,
+  "freq_shift": 0,
+  "in_channels": 4,
+  "layers_per_block": 2,
+  "mid_block_only_cross_attention": null,
+  "mid_block_scale_factor": 1,
+  "mid_block_type": "UNetMidBlock2DCrossAttn",
+  "norm_eps": 1e-05,
+  "norm_num_groups": 32,
+  "num_attention_heads": null,
+  "num_class_embeds": null,
+  "only_cross_attention": false,
+  "out_channels": 4,
+  "projection_class_embeddings_input_dim": 2816,
+  "resnet_out_scale_factor": 1.0,
+  "resnet_skip_time_act": false,
+  "resnet_time_scale_shift": "default",
+  "sample_size": 128,
+  "time_cond_proj_dim": null,
+  "time_embedding_act_fn": null,
+  "time_embedding_dim": null,
+  "time_embedding_type": "positional",
+  "timestep_post_act": null,
+  "transformer_layers_per_block": [
+    1,
+    2,
+    10
+  ],
+  "up_block_types": [
+    "CrossAttnUpBlock2D",
+    "CrossAttnUpBlock2D",
+    "UpBlock2D"
+  ],
+  "upcast_attention": null,
+  "use_linear_projection": true
+}

backend/huggingface/CabalResearch/Mugen/vae/config.json

@@ -0,0 +1,40 @@
+{
+  "_class_name": "AutoencoderKLFlux2",
+  "_diffusers_version": "0.37.0.dev0",
+  "_name_or_path": "black-forest-labs/FLUX.2-dev",
+  "act_fn": "silu",
+  "batch_norm_eps": 0.0001,
+  "batch_norm_momentum": 0.1,
+  "block_out_channels": [
+    128,
+    256,
+    512,
+    512
+  ],
+  "down_block_types": [
+    "DownEncoderBlock2D",
+    "DownEncoderBlock2D",
+    "DownEncoderBlock2D",
+    "DownEncoderBlock2D"
+  ],
+  "force_upcast": true,
+  "in_channels": 3,
+  "latent_channels": 32,
+  "layers_per_block": 2,
+  "mid_block_add_attention": true,
+  "norm_num_groups": 32,
+  "out_channels": 3,
+  "patch_size": [
+    2,
+    2
+  ],
+  "sample_size": 1024,
+  "up_block_types": [
+    "UpDecoderBlock2D",
+    "UpDecoderBlock2D",
+    "UpDecoderBlock2D",
+    "UpDecoderBlock2D"
+  ],
+  "use_post_quant_conv": true,
+  "use_quant_conv": true
+}

backend/loader.py

@@ -14,6 +14,7 @@ from backend.diffusion_engine.chroma import Chroma
 from backend.diffusion_engine.flux import Flux
 from backend.diffusion_engine.flux2 import Flux2
 from backend.diffusion_engine.lumina import Lumina2
+from backend.diffusion_engine.mugen import Mugen
 from backend.diffusion_engine.qwen import QwenImage
 from backend.diffusion_engine.sd15 import StableDiffusion
 from backend.diffusion_engine.sdxl import StableDiffusionXL, StableDiffusionXLRefiner
@@ -35,7 +36,7 @@ from backend.utils import (
 )
 from modules_forge.packages.comfy.utils import convert_diffusers_mmdit
 
-possible_models = [StableDiffusion, StableDiffusionXLRefiner, StableDiffusionXL, Chroma, Flux, Flux2, Wan, QwenImage, Lumina2, ZImage, Anima]
+possible_models = [StableDiffusion, StableDiffusionXLRefiner, StableDiffusionXL, Mugen, Chroma, Flux, Flux2, Wan, QwenImage, Lumina2, ZImage, Anima]
 
 logger = logging.getLogger("loader")
 setup_logger(logger)

backend/nn/vae.py

@@ -355,6 +355,8 @@ class AutoencoderKLFlux2(IntegratedAutoencoderKL):
         )
         self.bn.eval()
 
+        self.mugen = False  # 32 <-> 128
+
     def encode(self, x):
         z = super().encode(x)
 
@@ -373,9 +375,11 @@ class AutoencoderKLFlux2(IntegratedAutoencoderKL):
             eps=self.bn_eps,
         )
 
+        z = self.postprocess_encode(z)
         return z
 
     def decode(self, z):
+        z = self.preprocess_decode(z)
         s = torch.sqrt(memory_management.cast_to(self.bn.running_var.view(1, -1, 1, 1), dtype=z.dtype, device=z.device) + self.bn_eps)
         m = memory_management.cast_to(self.bn.running_mean.view(1, -1, 1, 1), dtype=z.dtype, device=z.device)
         z = z * s + m
@@ -393,3 +397,34 @@ class AutoencoderKLFlux2(IntegratedAutoencoderKL):
 
     def process_out(self, latent):
         return latent
+
+    def preprocess_decode(self, latent: torch.Tensor):
+        packed_channels: int = latent.size(1)
+        latent_channels: int = 128
+        scale_factor: int = 2
+
+        if self.mugen:
+            h = latent.shape[-2]
+            w = latent.shape[-1]
+            if h % scale_factor != 0 or w % scale_factor != 0:
+                pad_h = (scale_factor - (h % scale_factor)) % scale_factor
+                pad_w = (scale_factor - (w % scale_factor)) % scale_factor
+                latent = torch.nn.functional.pad(latent, (0, pad_w, 0, pad_h))
+                h = latent.shape[-2]
+                w = latent.shape[-1]
+            latent = latent.reshape(latent.shape[0], packed_channels, h // scale_factor, scale_factor, w // scale_factor, scale_factor)
+            latent = latent.permute(0, 1, 3, 5, 2, 4).reshape(latent.shape[0], latent_channels, h // scale_factor, w // scale_factor)
+
+        return latent
+
+    def postprocess_encode(self, latent: torch.Tensor):
+        packed_channels: int = 32
+        scale_factor: int = 2
+
+        if self.mugen:
+            h = latent.shape[-2]
+            w = latent.shape[-1]
+            latent = latent.reshape(latent.shape[0], packed_channels, scale_factor, scale_factor, h, w)
+            latent = latent.permute(0, 1, 4, 2, 5, 3).reshape(latent.shape[0], packed_channels, h * scale_factor, w * scale_factor)
+
+        return latent

backend/patcher/vae.py

@@ -121,7 +121,7 @@ def tiled_scale(samples, function, tile_x=64, tile_y=64, overlap=8, upscale_amou
 
 
 class VAE:
-    def __init__(self, model=None, device=None, dtype=None, no_init=False, *, is_wan=False, is_flux2=False):
+    def __init__(self, model=None, device=None, dtype=None, no_init=False, *, is_wan=False, is_flux2=False, is_mugen=False):
         if no_init:
             return
 
@@ -135,10 +135,10 @@ class VAE:
             self.memory_used_encode = lambda shape, dtype: (1767 * shape[2] * shape[3]) * memory_management.dtype_size(dtype)
             self.memory_used_decode = lambda shape, dtype: (2178 * shape[2] * shape[3] * 64) * memory_management.dtype_size(dtype)
 
-            if is_flux2:
+            if is_flux2 or is_mugen:
                 self.upscale_ratio = 16
                 self.downscale_ratio = 16
-                self.latent_channels = 128
+                self.latent_channels = 32 if is_mugen else 128
                 self.memory_used_decode = lambda shape, dtype: (2178 * shape[2] * shape[3] * 64) * memory_management.dtype_size(dtype) * 4.0
 
         else:
@@ -153,6 +153,8 @@ class VAE:
 
         self.output_channels = 3
         self.first_stage_model = model.eval()
+        if is_mugen:
+            self.first_stage_model.mugen = True
 
         self.device = device or memory_management.vae_device()
         offload_device = memory_management.vae_offload_device()

modules_forge/packages/huggingface_guess/latent.py

@@ -162,3 +162,10 @@ class Flux2(LatentFormat):
 
     def process_out(self, latent):
         return latent
+
+
+class SDXL_Flux2(Flux2):
+    def __init__(self):
+        super().__init__()
+        self.latent_rgb_factors_reshape = None
+        self.latent_channels = 32

modules_forge/packages/huggingface_guess/model_list.py

@@ -57,7 +57,7 @@ class BASE:
         return {}
 
     def inpaint_model(self):
-        return self.unet_config.get("in_channels", -1) > 4
+        return False
 
     def __init__(self, unet_config):
         self.unet_config = unet_config.copy()
@@ -115,6 +115,9 @@ class SD15(BASE):
     latent_format = latent.SD15
     memory_usage_factor = 1.0
 
+    def inpaint_model(self):
+        return self.unet_config.get("in_channels", -1) > 4
+
     def process_clip_state_dict(self, state_dict):
         k = list(state_dict.keys())
         for x in k:
@@ -179,6 +182,8 @@ class SDXL(BASE):
 
     unet_config = {
         "model_channels": 320,
+        "in_channels": 4,
+        "out_channels": 4,
         "use_linear_in_transformer": True,
         "transformer_depth": [0, 0, 2, 2, 10, 10],
         "context_dim": 2048,
@@ -189,6 +194,9 @@ class SDXL(BASE):
     latent_format = latent.SDXL
     memory_usage_factor = 0.8
 
+    def inpaint_model(self):
+        return self.unet_config.get("in_channels", -1) > 4
+
     def model_type(self, state_dict: dict):
         if "v_pred" in state_dict:
             return ModelType.V_PREDICTION
@@ -225,6 +233,23 @@ class SDXL(BASE):
         return {"clip_l": "text_encoder", "clip_g": "text_encoder_2"}
 
 
+class Mugen(SDXL):
+    huggingface_repo = "CabalResearch/Mugen"
+
+    unet_config = dict(SDXL.unet_config, in_channels=32, out_channels=32)
+
+    sampling_settings = {
+        "shift": 12.0,
+    }
+
+    latent_format = latent.SDXL_Flux2
+
+    vae_key_prefix = ["vae.", "first_stage_model."]
+
+    def inpaint_model(self):
+        return False
+
+
 class Flux(BASE):
     huggingface_repo = "black-forest-labs/FLUX.1-dev"
 
@@ -537,6 +562,7 @@ class QwenImage(BASE):
 models = [
     SD15,
     SDXL,
+    Mugen,
     SDXLRefiner,
     Flux,
     FluxSchnell,