DEV Community

AI-KSK
AI-KSK

Posted on Originally published at youtube.com

MiniMax H3 Acceleration in ComfyUI: Eleven Graphs Across Model Tier, LoRA, Attention and Cache

#comfyui #ai #video #workflows

MiniMax H3 acceleration is not one switch. The episode this set comes from measures four layers — model precision, LoRA, attention, and cache — and the eleven graphs here are the actual experiment matrix. Three VRAM tiers of image-to-video (24/48/80 GB), three tiers of reference-to-video, a W4A8 vs int8 comparison, Larry Turbo, LightX2V Turbo, Larry + SageAttention, and the Larry + attention + cache "full blood" build.

No inference was run for this release. There is no timing, VRAM number or quality comparison here, and the episode's "6x acceleration" claim is the author's description — not something I measured.

One attribution note: the graphs' notes credit the official Comfy-Org/MiniMax-H3 model files. Preserved as-is; this article documents the templates rather than claiming to author them.

# Route Nodes / links Saved sampler
01 I2V 24 GB (bf16) 24 / 29 euler + beta, 6 steps
02 W4A8 vs int8 comparison 62 / 78 res_multistep + simple, 20 steps
03 Larry + SageAttention (4 branches) 91 / 111 euler + beta, 6 steps
04 Larry Turbo 21 / 27 euler + beta, 6 steps
05 LightX2V Turbo 21 / 27 er_sde + simple, 4 steps
06 R2V 24 GB (bf16) 27 / 34 euler + beta, 6 steps
07 R2V 48 GB (int8) 27 / 34 euler + beta, 6 steps
08 I2V 48 GB (int8) 24 / 29 euler + beta, 6 steps
09 Larry + attention + cache 71 / 86 euler + beta, 6 steps
10 R2V 80 GB (pruned int8) 27 / 34 euler + beta, 6 steps
11 I2V 80 GB (pruned int8) 24 / 29 euler + beta, 6 steps

What the matrix actually varies

Every turbo-LoRA graph loads minimax_h3_turbo_v4_step600_ema_pruned_comfyui.safetensors at strength 1 and samples with euler + beta at 6 steps. The LightX2V graph breaks that pattern deliberately: minimax_h3_fl2v_lightx2v_turbo_4step_v0.1_comfy at 0.75, er_sde, 4 steps. If you compare wall-clock between those two files and call it a LoRA comparison, you are comparing two different recipes at once.

The tier files differ only in quantization:

  • 24 GB: fl2va_bf16 / ref2va_bf16 with bf16 text encoder and fp16 video VAE.
  • 48 GB: fl2va_int8_convrot / ref2va_int8_convrot with nvfp4-awq encoder and int8 video VAE.
  • 80 GB: fl2va_pruned_int8_convrot / ref2va_pruned_int8_convrot, same encoder/VAE tier.

Image-to-video files take one LoadImage. Reference files take LoadImage + LoadVideo + LoadAudio. All of them save to video/MiniMax_H3.

The comparison rigs, and why they are easy to misread

Three files are not single pipelines:

  • 02 INT8 W4A8 (62 / 78) holds three parallel branches: one W4A8 mixed model and two int8-convrot branches, all keeping the full 20-step res_multistep recipe and no LoRA. That makes it the "no acceleration tricks, just quantization" reference point — and it is also the only file in the set where the sampler is still 20 steps.
  • 03 Larry + SageAttention (91 / 111) has four branches on the same int8 model and turbo LoRA, differing in attention/cache patches.
  • 09 "full blood" (71 / 86) has three branches with cache patches layered on the int8 model.

All three save every branch to the same output prefix. Before you compare anything, read which branch is actually wired to the save node — otherwise you will be timing one branch and reading another's output.

Picking a starting point

If your goal is "run one graph today", start with 04 Larry Turbo (21 nodes, one model, one LoRA, six steps) or the tier file that fits your GPU. If your goal is to reproduce the experiment, run the tier files first to establish a baseline, then the comparison rigs.

Two practical checks before queueing:

  • Confirm the LoRA strength (1.0 for turbo-v4, 0.75 for LightX2V).
  • Confirm the model variant (bf16 / int8_convrot / pruned_int8_convrot) — the filenames differ by a suffix that is easy to miss, and mixing tiers will produce a result that is not comparable.

What this is not

Not a benchmark, not a speed test, and not a license grant. The video's "6x sweet spot" framing is the author's description; this article does not claim to have measured an acceleration factor. The graphs' notes credit the official Comfy-Org/MiniMax-H3 model files. The files themselves grant no rights to the H3 weights, the LoRAs, the encoder or the VAEs.

Getting the workflows

Published by this channel. The graphs are free; the model weights, LoRAs, encoder, VAEs and example media are not bundled.

Verify workflow, node, model, LoRA, encoder, VAE, input-media and cover rights separately. No blanket commercial or redistribution license is asserted for the graphs or the example media.

Top comments (0)