Minimax-H3

PlagueKind - Minimax H3 - Workflow - Ease of use

Help me get a 5090

<div style="display: flex; gap: 10px;"> <img src="https://cdn-uploads.huggingface.co/production/uploads/699b9e22df8352ba379115d6/YDecfaadDgVO29LblZj0L.png" width="480"/> </div> <video controls width="50%"> <source src="https://cdn-uploads.huggingface.co/production/uploads/699b9e22df8352ba379115d6/-2Ytq1lSVM6KCW2AiKPj0.mp4"> Your browser does not support the video tag. </video>

PlagueKind - Minimax H3 - LTX Workflow - Ease of use - Eros | Sulphur Compatible | FaceID

Minimax H3

V11 Notes

• UPDATE NODEPACK TO V1.5.6 AND COMFYUI TO NEWEST VERSION, and requirements.
• Added Refmod Support, create in subgraph.
• Got rid of fantastic builder kept only luisachaotica.
• Fixed layout issues and removed unwanted dependencies.
• Dropped turbo, it was ruining video and audio. default WF settings will get you 4 skips for "9" steps.
• Added my H3 Combo node that does all modes.
• Overall should be better prompt adherence, faster, and less overhead with reduced peak vram.
• Use reference map to know what the LM is calling everything.
• If you don't like no turbo, just use your favorite settings from before. Shift node in Sampler subgraph for people who want to keep using turbo.

Support

If you find this project useful and want to support development:

Monero (XMR):

865BrcfWLdwELwuq5faV1uVTbh93zVK6AUYLY2c3mX6sFfAGRfS6axe1kBTYYKuM7ccN7zBZDAZvnT7E4NKmUazySdbpc7p

V9 Notes

• UPDATE NODEPACK TO V1.4.7 AND COMFYUI TO NEWEST VERSION. BEFORE EVEN DOWNLOADING THE WF.

• New mode for sparse, comfy engine, better quality and faster.

• Added Int8 Kernel.

• I set 0.7 sparsity because it's not much slower on comfy engine, mess with it.

• Added H3 cache, default is safe.

• Added Bypass for R2V VAE.

• Upscaling is still WIP, mess with it.

V8 Notes

• Update nodepack to v1.4.3 first

• Added my turbo lora, functional strength up to 5.0

• Test default settings, they should be good for most prompts.

• If you don't like mine on it's own, try blending the disabled fast lora with it, looks really good.

• Res multi step simple also works well.

• Added tiled upscaling, thanks to lora daddy.

V7 Notes

• SLA has reference protection now, speeds up R2V signficantly, also fixed slowdown from max ref size. Turning it off gives a speed boost for barely any loss.
• Added blocksize 32 to improve quality. if you want the old slightly faster behavior enter subgraph, set blocksize 64.
• 768 Dareties turbo added combined with the old Dareties for 6 steps.
• Use Seeds_2 + Sgm_Uniform for best quality. (slower, but 4 steps looks great). Old dareties 1str, or 768 at 0.75str
• If using single audio sources like one voice, use shift 3 audio for cleaner voice. if there is music underneath and sound effects, try audio shift 5

V6 Notes

• SLA node overhaul, vast improvement of control and quality.
• Added ER_SDE ODE override.

V5.6 Notes

• Unified resize now has Aspect ratio selector

• Changed default sampler to sa_solver

• I needed to make this because changing unified resize breaks old workflows. update nodepack before loading the workflow

V5.5 Notes

• Major'ish Overhaul of the layout, and wiring.

• Disabling Sparse falls back to comfy attention.

• Sparse Density Notes:

  • 0.80-0.85 - "match pytorch quality"

  • 0.9 - Default - Minimal Degredation 15-20% faster than 0.85

  • 0.95 - HUGE boost for longer hi res generations, can introduce minor degredation

V5 Notes

• I've added my Sparse attention node. No other speedups required. up to 2.5x speedup.

• Switched to dareties lora. Use with shift 8/3 0.8 strength 6-8 steps. this lora works great for all modes.

• Memory Chunking, if you go to high with sparse, it causes slow down. for me beyond 2-4 caused slowdown on 16gb vram.

• Don't mess with the sparse settings on the subgraph unless you have to.

• Make sure to update my node pack to atleast 1.3.1

• Make sure you're on the newest pytorch+cu combo supported.

V4 Notes

• Sage begone, welcome comfy attn. update comfyui, sage without sage issues.

• Switched to 4 step 768 lora. Use with shift 6/3 0.8 strength 6-8 steps, or 1.0 with 4-6 steps.

• Added hybrid model so you don't need to swap anymore, start with 25. 30 higher quality slightly less ref strength.

• When ref swapping a video use er_sde/beta

• Added original audio mode, avoids degradation and vae decode / Allows you to make music videos etc. it's in the reference subgraph.

• Added minimax mode and mode selection to lora stack, so update node pack.

V3.1 Notes

• Sage begone, welcome comfy attn. update comfyui, sage without sage issues.

• Switched to 8 step new lora, these 2 new things combined behave very close to non turbo non sage.

V3 Notes

• Added R2V mode.

V2.1 Notes

• Added preview node.

V2 Notes

• Added int8 convrot video vae

• Kijai resized turbo lora, best at 0.6-0.7 strength.

• Memory chunks return. huge speed up for me. 4 was pretty much the cap for 16gb vram.

---

Requirements

• ComfyUI-PlagueKind-Nodes
• ComfyUI-KJNodes
• Rgthree-comfy
• Nvidia_RTX_Nodes_ComfyUI
• ComfyUI-VideoHelperSuite
• ComfyUI-H3-Motion-Context-MultiRef
• ComfyUI-MiniMaxH3Mod

---

Diffusion Model

minimax_h3_hybrid_fl2va_ref2va_b25-49-int8.safetensors

minimax_h3_hybrid_fl2va_ref2va_b30-49-int8.safetensors

VAE

minimax_h3_audio_vae_fp32.safetensors

minimax_h3_video_vae_int8_convrot.safetensors

vae_approx/taeh3.safetensors

Text Encoder

qwen3vl_32b_minimax_h3_int8_convrot.safetensors

Frame Interpolation

film_net_fp16.safetensors

Latent Upscaler

minimax_h3_latent_upscaler_3d_bf16.safetensors

Optional Turbo Loras

H3-PK-Parasyte-Turbo.safetensors

fasth3_6step.safetensors

supporting file required. make the folder models/h3_adaln

h3_silu_temb_grid.safetensors

---

V1.5 Notes

• Added the option for T2V.

• When using T2V, set scale mode of Target Dimension node to Dimensions (WxH) or you will get a square video.

• Added FSR RCAS sharpening, 0.3 strength looks natural.

• FFLF is unreliable at res above 640, don't know why, it's like this on default WF as well.

• Target Dimension node controls the generation size only, refs are passed elsewhere.

• Don't touch the other hidden resize nodes as they only control ref size and are set to max.

• Euler + Linear Quadratic seems to be the best combo with between 15-20 steps.

• I didn't add easy cache because it ruins the quality noticeably,

• If you have time or a powerful card, try heunpp2 sampler. It's phenomenal.