zlei6/ComfyUI
1
0
Fork 0
mirror of https://github.com/Comfy-Org/ComfyUI.git synced 2026-08-25 02:20:18 +08:00
Code Issues Packages Projects Releases Wiki Activity

feat: Support MiniMax-H3 (CORE-375) (#15224)

Browse Source
This commit is contained in:
Jukka Seppänen
2026-08-03 05:28:29 +03:00
committed by GitHub
parent b53e247c94
commit 57500fc5bc
16 changed files with 2599 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

52
comfy/model_base.py
View File

@@ -21,6 +21,7 @@ import comfy.ldm.hunyuan3dv2_1.hunyuandit
import torch
import logging
import comfy.ldm.lightricks.av_model
import comfy.ldm.minimax.model
import comfy.ldm.lightricks.symmetric_patchifier
import comfy.context_windows
from comfy.ldm.modules.diffusionmodules.openaimodel import UNetModel, Timestep
@@ -2063,6 +2064,57 @@ class Hunyuan3Dv2_1(BaseModel):
out['guidance'] = comfy.conds.CONDRegular(torch.FloatTensor([guidance]))
return out
class MiniMaxH3(BaseModel):
def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.minimax.model.MiniMaxH3Model)
def extra_conds(self, **kwargs):
out = super().extra_conds(**kwargs)
cross_attn = kwargs.get("cross_attn", None)
if cross_attn is not None:
# run condition_proj + token refiner once per sampling instead of per step
cross_attn = self.diffusion_model.preprocess_text_embeds(
cross_attn.to(device=kwargs["device"], dtype=self.get_dtype_inference()))
out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
latent_shapes = kwargs.get("latent_shapes", None)
if latent_shapes is not None:
out['latent_shapes'] = comfy.conds.CONDConstant(latent_shapes)
# Everything H3-specific rides in one dict so _apply_model's dtype cast
# (which would flatten fp32 cond latents and long tags to bf16) skips it.
payload = {}
tags = kwargs.get("minimax_token_tags", None)
if tags is not None:
payload["text_token_tags"] = tags
keyframes = kwargs.get("minimax_keyframes", None)
if keyframes is not None:
payload["keyframes"] = keyframes
payload["frame_count"] = kwargs.get("minimax_frame_count", None)
payload["cond_video_latents"] = [kf["latent"] for kf in keyframes]
refs = kwargs.get("minimax_refs", None)
if refs is not None:
payload["refs"] = refs
payload["cond_video_latents"] = [r["latent"] for r in refs if "latent" in r]
payload["cond_audio_latents"] = [r["audio_latent"] for r in refs if r.get("audio_latent") is not None]
if kwargs.get("minimax_visual_cond_noise_aug", None) is not None:
payload["visual_cond_noise_aug"] = kwargs["minimax_visual_cond_noise_aug"]
if kwargs.get("minimax_audio_cond_noise_aug", None) is not None:
payload["audio_cond_noise_aug"] = kwargs["minimax_audio_cond_noise_aug"]
payload["seed"] = kwargs.get("seed", 0)
if cross_attn is not None and latent_shapes is not None and len(latent_shapes) > 1:
# packed layout built once per sampling run, h/w rounded up to the DiT's 2x2 patch
vs = latent_shapes[0]
payload["layout"] = comfy.ldm.minimax.model.PackedLayout(
cross_attn.shape[1], vs[2], (vs[3] + 1) // 2 * 2, (vs[4] + 1) // 2 * 2,
latent_shapes[1][-1], keyframes=payload.get("keyframes"),
refs=payload.get("refs"), frame_count=payload.get("frame_count"))
out['minimax_payload'] = comfy.conds.CONDConstant(payload)
return out
def scale_latent_inpaint(self, sigma, noise, latent_image, **kwargs):
return latent_image
class TripoSplat(BaseModel):
def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.triposplat.model.LatentSeqMMFlowModel)