MiniMax H3 ComfyUI Kitchen Attention Mode - For Patreon Supporters (Patreon)

Hey everyone.

If you watched the Kitchen Attention vs SageAttention video, you saw the side-by-side comparison at 20 and 30 steps.

https://www.patreon.com/aifuturetech/posts/comfyui-kitchen-167312486

This post has the complete workflow with both attention modes plus two optional upscale paths to push your output to 720p or 4K. And some example text prompt for test with Minimax H3.

1. MiniMax-H3 TXT2VID IMG2VID (Full)- 20260818.json

This is the full production workflow. The basic version from the video is a stripped-down test case — this one has the complete pipeline with everything wired up for real use.

1. MiniMax H3 Text-to-Video / Image-to-Video (Full)

....................................................

File: MiniMax-H3 TXT2VID IMG2VID (Full)- 20260818.json

The core pipeline generates video from a text prompt (or image + prompt for img2vid), then optionally runs it through two upscale stages. The workflow is built around Set/Get nodes for routing, so you can enable or disable sections without reconnecting anything.

Pipeline:

- Load Models: MiniMax H3 UNet (int8 quantized) + Qwen3 VL 32B text encoder (int8) + H3 video VAE + H3 audio VAE

- Text Encoding: CLIP text encode with positive and negative prompts via Qwen3 VL

- Attention Mode: Choose ONE per run — either ComfyUI Kitchen Attention or SageAttention (ModelAttentionBackend node). Bypass the other.

- Turbo LoRA: minimax_h3_fl2v_turbo_4step (bypassed — not used in this workflow)

- Sampling: MiniMaxH3SigmaShift + BasicGuider + BasicScheduler + SamplerCustomAdvanced

- Decode: VAEDecode for video frames, VAEDecodeAudio for audio track

- Upscale Stage 1 (optional): SeedVR2 720p upscale — loads the SeedVR2 DiT and VAE models, upscales the video frames

- Upscale Stage 2 (optional): LTX 2.5 Latent Spatial Upscaler 2x — takes the SeedVR2 output to 1080p in latent space

- Alternative Upscale Path: RTX Video Super Resolution + Frame Interpolation (RIFE) — connects directly from H3 output to 4K with smooth frame interpolation

- Output: VHS_VideoCombine for final video export with audio passthrough

How the routing works:

The workflow uses Get/Set nodes throughout. This means:

- Choose your attention mode: Set the Kitchen Attention group to "enable" and SageAttention to "bypass", or vice versa. Never run both at the same time.

- Choose your upscale path: If you have the VRAM, enable the SeedVR2 + LTX 2.5 chain for 1080p. If you are tight on VRAM, bypass SeedVR2 and connect the H3 output directly to the RTX Super Resolution + Frame Interpolation group for 4K upscaling with frame smoothing.

- The Get/SET nodes handle all the routing — no need to draw node lines manually. Just toggle the groups on and off.

Key settings:

- UNet: minimax_h3_fl2va_int8_convrot.safetensors (int8 quantized for lower VRAM)

- Text Encoder: qwen3vl_32b_minimax_h3_int8_convrot.comfy.safetensors (int8)

- Video VAE: minimax_h3_video_vae_fp16.safetensors

- Audio VAE: minimax_h3_audio_vae_fp32.safetensors

- Turbo LoRA: minimax_h3_fl2v_turbo_4step_v1.1_768p_comfyui_bf16.safetensors (bypassed — not used in this workflow)

- SeedVR2 DiT: via SeedVR2LoadDiTModel (custom node)

- SeedVR2 VAE: via SeedVR2LoadVAEModel

- LTX 2.5 Upscaler: ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors

- LTX 2.5 Transformer: ltx-2.5-22b-dev-transformer-bf16.safetensors

- LTX 2.5 Text Encoder: gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot.safetensors

- LTX 2.5 VAE: ltx-2.5-audio-vae-bf16.safetensors

- Spatial Upscaler: ltx-2.3-spatial-upscaler-x2-1.1.safetensors

- RTX Super Res: RTXVideoSuperResolution node (requires NVIDIA RTX GPU)

- Frame Interpolation: film_net_fp16.safetensors (RIFE model)

----------------------------------------

WHAT YOU NEED

----------------------------------------

Model files (exact paths):

diffusion_models/

minimax-h3/minimax_h3_fl2va_int8_convrot.safetensors

text_encoders/

minimax-h3/qwen3vl_32b_minimax_h3_int8_convrot.comfy.safetensors

ltx-2.5/gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot.safetensors

ViT-B-32.pt (CLIP for frame interpolation)

vae/

minimax-h3/minimax_h3_video_vae_fp16.safetensors

minimax-h3/minimax_h3_audio_vae_fp32.safetensors

ltx-2.5/ltx-2.5-audio-vae-bf16.safetensors

loras/

minimax-h3/minimax_h3_fl2v_turbo_4step_v1.1_768p_comfyui_bf16.safetensors (bypassed)

upscale_models/

ltx-2.5/ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors

ltx-2.3/ltx-2.3-spatial-upscaler-x2-1.1.safetensors

film_net_fp16.safetensors (RIFE)

diffusion_models/ (for LTX 2.5 upscaler)

ltx-2.5/ltx-2.5-22b-dev-transformer-bf16.safetensors

Download MiniMax H3 models from:

https://huggingface.co/Comfy-Org/MiniMax-H3

Download LTX 2.5 models from:

https://huggingface.co/Comfy-Org/LTX-Video-2.5

Custom nodes:

- ComfyUI-MiniMax-H3 — MiniMaxH3ImageToVideo, MiniMaxH3SigmaShift

- seedvr2_videoupscaler — SeedVR2LoadDiTModel, SeedVR2LoadVAEModel, SeedVR2VideoUpscaler

- ComfyUI-LTXVideo — LTX latent spatial upscaler

- ComfyUI-Frame-Interpolation — FrameInterpolate, FrameInterpolationModelLoader

- ComfyUI-RTX-Video-Super-Resolution — RTXVideoSuperResolution

- ComfyUI-KJNodes — PathchSageAttentionKJ

- ComfyUI-ModelAttentionBackend — ModelAttentionBackend (Kitchen/Sage toggle)

- rgthree — Power Lora Loader

- ComfyUI-VideoHelperSuite — VHS_VideoCombine, video output

GPU requirements:

- Basic H3 generation (no upscale): 8GB VRAM minimum, 12GB+ recommended

- SeedVR2 + LTX 2.5 upscale to 1080p: 16GB+ VRAM recommended

- RTX Super Resolution + Frame Interpolation: requires NVIDIA RTX GPU (RTX 20-series or newer) with drivers that support RTX Video Super Resolution

----------------------------------------

QUICK START

----------------------------------------

1. Import the workflow JSON into ComfyUI

2. Check that all model paths match your local folders — the workflow uses forward-slash paths relative to ComfyUI directories

3. Choose your attention mode: Enable either the "Patches W/ ComfyUI Kitchen Att." group or the "Patches W/ SageAtt" group. Set the other to bypass.

4. Write your text prompt in the text prompt node

5. Set your resolution and duration in the User Input group

6. Optional: Enable the SeedVR2 + LTX 2.5 upscale group if you have 16GB+ VRAM

7. Alternative: Enable the RTX Super Resolution + Frame Interpolation group for 4K output on RTX GPUs

8. Queue the workflow and check outputs at each stage (the workflow has 5 video output nodes so you can preview at different pipeline stages)

----------------------------------------

TIPS

----------------------------------------

1. Always run one attention mode at a time. The ModelAttentionBackend node lets you switch between Kitchen and Sage — set one to enable and the other to bypass. Running both at the same time wastes VRAM and can cause conflicts.

2. For the upscale path, SeedVR2 gives you the best quality at 720p, and the LTX 2.5 spatial upscaler pushes it to 1080p. If your GPU has 16GB+ VRAM, this is the path to use.

3. If you are on 12GB or less, skip SeedVR2 and use the RTX Super Resolution + Frame Interpolation path instead. It connects directly from the H3 output and uses less VRAM. The RTX upscaler requires an NVIDIA RTX GPU, but if you have one, the quality is solid.

4. The GET/SET node routing means you do not need to manually reconnect node lines when switching between upscale paths. Just toggle the group enable/bypass. This is the main advantage of the full workflow over the basic version.

5. Frame interpolation with RIFE (film_net_fp16) smooths out the motion between frames. Use it when your output looks choppy at low frame rates. It adds processing time but the difference is noticeable on slow-motion scenes.

6. The workflow has 5 video output nodes — use them to preview at different stages: raw H3 output, post-SeedVR2, post-LTX upscale, post-RTX, and final interpolated. This helps you decide which upscale level you actually need.

7. Start with the basic workflow from the video to test your prompts and settings, then switch to this full workflow when you are ready for the final render with upscaling.

----------------------------------------

COMMON ISSUES

----------------------------------------

SeedVR2 out of memory — SeedVR2 is VRAM hungry. Drop to 720p resolution or reduce the frame count. If you are on 12GB, use the RTX Super Resolution path instead.

LTX 2.5 upscaler producing blurry output — make sure you are using the correct spatial upscaler model (ltx-2.3-spatial-upscaler-x2-1.1.safetensors, not the raw LTX 2.5 checkpoint). The spatial upscaler is specifically trained for 2x latent space upscaling.

RTX Super Resolution not available — this requires an NVIDIA RTX GPU (20-series or newer) with up-to-date drivers. Check that RTX Video Super Resolution is enabled in NVIDIA Control Panel > Video > Adjust Video Image Settings.

Audio not playing in output — check that the audio VAE decode is connected properly. The workflow has separate VAE paths for video and audio. Make sure both are connected to the final VHS_VideoCombine node.

Attention mode not switching — the ModelAttentionBackend node must be set to either "ComfyUI Kitchen" or "SageAttention" in the dropdown. Make sure the node you are NOT using is set to bypass mode, not just disconnected.

Frame interpolation artifacts — if RIFE is producing ghosting or strange motion artifacts, reduce the interpolation multiplier. The default is usually 2x — try 1.5x if the motion looks unnatural.

----------------------------------------

All model files need to be downloaded separately from the links above before you run the workflow.

Questions? Drop them in the comments and I will help you sort it out.

Thanks for supporting the channel. You guys make these deep dives possible.

— Benji