Uploaded August 2026 | Updated November 2026, 1 day ago
I used MiniMax Music 3 in ComfyUI to create a complete song, then turned it into a full MiniMax H3 music video using the H3 Turbo 4-Step LoRA.
I tested three versions of the same song with the same lyrics but different music captions. Then I used H3 Ref2VA, character and environment references, a 1-second audio overlap for cleaner lip-sync cuts, the Spatial Physics LoRA, and a two-pass 2× latent upscale workflow.
In the video:
• MiniMax Music 3 workflow and prompting
• Simple vs detailed music captions
• H3 Ref2VA music-video workflow
• 1-second lip-sync overlap trick
• Character + environment consistency
• H3 Turbo 4-Step LoRA
• Physics LoRA at 0.25
• EasyCache lip-sync comparison
• H3 two-pass 2× upscale
## DOWNLOADS
**MiniMax Music 3**
Models:
huggingface.co/Comfy-Org/MiniMax-Music-3
Official model:
huggingface.co/MiniMaxAI/MiniMax-Music3
ComfyUI guide:
docs.comfy.org/tutorials/audio/minimax/minimax-music-3
Files used/available:
`minimax_music3_dit_fp32.safetensors`
`minimax_music3_dit_fp16.safetensors`
`minimax_music3_dit_int8_convrot.safetensors`
`minimax_music3_text_encoder_bf16.safetensors`
`minimax_music3_text_encoder_pruned_bf16.safetensors`
`minimax_music3_text_encoder_pruned_int8_convrot.safetensors`
`minimax_music3_dav.safetensors`
**MiniMax H3**
All ComfyUI H3 models:
huggingface.co/Comfy-Org/MiniMax-H3
H3 Ref2VA INT8:
huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/diffusion_models/minimax_h3_ref2va_pruned_int8_convrot.safetensors
Qwen3-VL text encoder:
huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors
Video VAE:
huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/vae/minimax_h3_video_vae_fp16.safetensors
Audio VAE:
huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/vae/minimax_h3_audio_vae_fp32.safetensors
Turbo 4-Step LoRA:
huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/loras/minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors
**Physics LoRA**
huggingface.co/Jojocodex/minimax-h3-wushu-action-lora
File used:
`wushu_spatial_physics_v2_pruned.safetensors`
**H3 2× Latent Upscaler**
Custom node:
github.com/Tr1dae/ComfyUI-MiniMaxH3_LatentUpscaler
Workflow-selected model:
`h3_clean_latent_upscaler_film_epoch200.safetensors`
**Useful Nodes / Workflows**
KJNodes:
github.com/kijai/ComfyUI-KJNodes
Official H3 ComfyUI workflow:
github.com/Comfy-Org/workflow_templates/blob/main/templates/video_minimax_h3_r2v.json
Workfllow
drive.google.com/file/d/1RFmU5wf1Xn_amuYSjiNvOs5LnU7eriB7/view?usp=sharing
drive.google.com/file/d/1NS8lhfDcjiXLrrUSMiaVy1tDXoYyC1W2/view?usp=sharing
For my final music-video clips, bypassing EasyCache gave me better lip-sync results. Test both settings with the same scene because your result may be different.
#MiniMaxMusic3 #MiniMaxH3 #ComfyUI
I used MiniMax Music 3 in ComfyUI to create a complete song, then turned it into a full MiniMax H3 music video using the H3 Turbo 4-Step LoRA.
I tested three versions of the same song with the same lyrics but different music captions. Then I used H3 Ref2VA, character and environment references, a 1-second audio overlap for cleaner lip-sync cuts, the Spatial Physics LoRA, and a two-pass 2× latent upscale workflow.
In the video:
• MiniMax Music 3 workflow and prompting
• Simple vs detailed music captions
• H3 Ref2VA music-video workflow
• 1-second lip-sync overlap trick
• Character + environment consistency
• H3 Turbo 4-Step LoRA
• Physics LoRA at 0.25
• EasyCache lip-sync comparison
• H3 two-pass 2× upscale
## DOWNLOADS
**MiniMax Music 3**
Models:
huggingface.co/Comfy-Org/MiniMax-Music-3
Official model:
huggingface.co/MiniMaxAI/MiniMax-Music3
ComfyUI guide:
docs.comfy.org/tutorials/audio/minimax/minimax-music-3
Files used/available:
`minimax_music3_dit_fp32.safetensors`
`minimax_music3_dit_fp16.safetensors`
`minimax_music3_dit_int8_convrot.safetensors`
`minimax_music3_text_encoder_bf16.safetensors`
`minimax_music3_text_encoder_pruned_bf16.safetensors`
`minimax_music3_text_encoder_pruned_int8_convrot.safetensors`
`minimax_music3_dav.safetensors`
**MiniMax H3**
All ComfyUI H3 models:
huggingface.co/Comfy-Org/MiniMax-H3
H3 Ref2VA INT8:
huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/diffusion_models/minimax_h3_ref2va_pruned_int8_convrot.safetensors
Qwen3-VL text encoder:
huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors
Video VAE:
huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/vae/minimax_h3_video_vae_fp16.safetensors
Audio VAE:
huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/vae/minimax_h3_audio_vae_fp32.safetensors
Turbo 4-Step LoRA:
huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/loras/minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors
**Physics LoRA**
huggingface.co/Jojocodex/minimax-h3-wushu-action-lora
File used:
`wushu_spatial_physics_v2_pruned.safetensors`
**H3 2× Latent Upscaler**
Custom node:
github.com/Tr1dae/ComfyUI-MiniMaxH3_LatentUpscaler
Workflow-selected model:
`h3_clean_latent_upscaler_film_epoch200.safetensors`
**Useful Nodes / Workflows**
KJNodes:
github.com/kijai/ComfyUI-KJNodes
Official H3 ComfyUI workflow:
github.com/Comfy-Org/workflow_templates/blob/main/templates/video_minimax_h3_r2v.json
Workfllow
drive.google.com/file/d/1RFmU5wf1Xn_amuYSjiNvOs5LnU7eriB7/view?usp=sharing
drive.google.com/file/d/1NS8lhfDcjiXLrrUSMiaVy1tDXoYyC1W2/view?usp=sharing
For my final music-video clips, bypassing EasyCache gave me better lip-sync results. Test both settings with the same scene because your result may be different.
#MiniMaxMusic3 #MiniMaxH3 #ComfyUI

![Qwen Image 2.1 Turbo Test: VAE, Samplers, PNGs & Image Editing
I tested Qwen Image 2.1 Turbo in ComfyUI against the full Qwen Image 2.1 model. With one sampler change, the 8-step Turbo result looked better than the full model at 25 steps in my tests, with the same seed and the same prompt.
I used the full Turbo BF16 model for every test. Here is what I tested:
- Normal 8 steps (KSampler, Euler, simple) vs the official 8-step schedule (ManualSigmas). Normal 8 steps gave more natural skin. The official schedule was better for hands, objects, and counting.
- Three VAEs: the official Qwen 2.1 VAE, the HDR VAE, and the Texture-Fix VAE
- The sampler setting that beat the full BF16 model
- Qwens prompt enhancer (the refine_prompt switch)
- Transparent PNGs in 8 steps, for new designs and for cutouts
- One VAE wire on TextEncodeQwenImage21 that decides if Qwen copies your image or reimagines it
- Resolution 0 vs 1024 for image editing
Settings that worked best for me:
Sampler: DPM++ SDE
Scheduler: linear quadratic
Steps: 8
CFG: 1
VAE: HDR VAE for Qwen Image 2.1
Please note: this is not Qwens official setting. Qwen says other schedules have not been tested with Turbo. DPM++ SDE also takes more time per step than Euler, so check the speed on your own GPU.
Official 8-step schedule for the ManualSigmas node (9 values):
[PASTE THE 9 SIGMA VALUES]
Transparent PNG wording (from Qwens model page):
Start your prompt with: This is an RGBA image with transparency.
End your prompt with: The image has an alpha channel and the background is transparent.
Official Qwen Image 2.1 Turbo sigma values: 1.0, 0.978453, 0.95418, 0.926626, 0.89508, 0.845148, 0.704534, 0.414568, 0.0
Downloads
Workflows: https://www.patreon.com/ComfyuiWorkflowBlog/posts/i-tested-qwen-2-171973198
Qwen Image 2.1 Turbo, full BF16 (around 14 GB): https://huggingface.co/Comfy-Org/Qwen-Image-2.1/tree/main/diffusion_models
Qwen Image 2.1 Turbo INT8 (7.26 GB):https://huggingface.co/Comfy-Org/Qwen-Image-2.1/tree/main/diffusion_models
Official Qwen Image 2.1 VAE:
https://huggingface.co/Comfy-Org/Qwen-Image-2.1/blob/main/vae/qwen_image_2.1_vae_bf16.safetensors
Lora: https://huggingface.co/Comfy-Org/Qwen-Image-2.1/tree/main/loras
HDR VAE for Qwen Image 2.1 (community):
https://civitai.com/models/2980907
Texture-Fix VAE for Qwen Image 2.1 (community, by madebyollin):
https://huggingface.co/madebyollin/texture-fix-vae-for-qwen-image-2.1
Where to put the files
Turbo model: ComfyUI/models/diffusion_models
Turbo LoRA: ComfyUI/models/loras
Try the DPM++ SDE setting on your own images, and tell me in the comments if it worked for you too.
#QwenImage #ComfyUI #AIImageGeneration Qwen Image 2.1 Turbo Test: VAE, Samplers, PNGs & Image Editing](https://i.ytimg.com/vi/WFUyVJIi5M0/mqdefault.jpg)


![Wan 2.1 LongCat Avatar Comfy Ui (Text to Video Hack) with Unlimited Length
I found a hidden switch in the Wan 2.1 LongCat Avatar model that changes everything. Most people think this workflow is just for extending avatar videos. But if you change one simple number, it stops looking at your image and turns into a full Text-to-Video generator.
In this video, I am going to show you exactly how to use this Overlap 0 Hack. I will also explain why I switched to 93 frames instead of 81, and the specific settings you need to fix the big eyes bug that everyone is complaining about.
If you have a 12 GB card or even lower, do not worry. This guide covers the FP8 version that runs smoothly on almost any modern GPU.
Song: Cartoon, Jéja - On & On (feat. Daniel Levi) [NCS Release]
Music provided by NoCopyrightSounds
Free Download/Stream: http://ncs.io/onandon
Watch: http://youtu.be/K4DyBUG242c
Downloads You Need
Main Model: LongCat_TI2V_comfy_fp8_e4m3fn_scaled_KJ.safetensors (Use the FP8 version for speed)
https://huggingface.co/Kijai/LongCat-Video_comfy/tree/main/Avatar
Speed LoRA: LongCat_distill_lora_alpha64_bf16.safetensors (Do not skip this or it will be slow)
https://huggingface.co/Kijai/LongCat-Video_comfy/blob/main/LongCat_distill_lora_alpha64_bf16.safetensors
Support Files: Wan 2.1 VAE, UMT5 XXL FP8 Text Encoder,
Workflow: https://aistudynow.com/longcat-avatar-comfyui-workflow/
Key Settings to Remember
For Avatar (Image to Video): Set Overlap to 14 or 16.
For New Character (Text to Video): Set Overlap to 0.
To Fix Big Eyes: Set Guidance (CFG) to 1.0 exactly. Wan 2.1 LongCat Avatar Comfy Ui (Text to Video Hack) with Unlimited Length](https://i.ytimg.com/vi/WYwjAumr-U8/mqdefault.jpg)





