README.md
7.3 KB · 153 lines · markdown Raw
1 ---
2 license: apache-2.0
3 base_model: Comfy-Org/MiniMax-H3
4 tags:
5 - text-to-video
6 - text-to-audio
7 - audio-video
8 - lora
9 - minimax-h3
10 - comfyui
11 pipeline_tag: text-to-video
12 ---
13
14 # MiniMax-H3 Turbo LoRA — few-step audio-video generation
15
16 A LoRA for [MiniMax-H3](https://huggingface.co/Comfy-Org/MiniMax-H3) that renders
17 joint **video + synchronized stereo audio** in as few as **4 sampling steps**
18 instead of the usual ~20 — a ~5× sampling speedup — and keeps getting better as
19 you add steps.
20
21 ## Which checkpoint — `v4` (step 600) or `v1` (850)?
22
23 For **most** work, use **`minimax_h3_turbo_v4_step600_ema.safetensors`**. It's the
24 strongest checkpoint we've released: much better static and small-motion shots,
25 markedly better micro-detail (faces, fingers, fine texture), and the
26 over-sharpening / plastic look of the earlier `v1` (~850) line is **fully
27 resolved**.
28
29 v4 introduced a **static-frame enhancement** — a big win for static and
30 small-motion content. The one trade-off shows up **only at 4 steps with large,
31 fast motion**, where v4 can produce **motion-smear / trailing ghosting** (we're
32 actively fixing this). Two things address it:
33
34 - **Use 6–8 steps.** This **largely removes the smear** and is where v4 looks its
35 best. v4 also tolerates higher step counts better than v1, which tends to
36 over-sharpen at high steps + strength 1.0.
37 - For the specific case of **4 steps *and* heavy motion**, the older **`v1` ~850**
38 checkpoint can still be the friendlier pick.
39
40 ```
41 Using 6–8 steps? ── yes ──► v4-600 (recommended)
42 │ no (4 steps)
43 ▼
44 Heavy / fast motion? ── no ──► v4-600 (recommended)
45 │ yes
46 ▼
47 v1-850 (friendlier at 4-step heavy motion)
48 ```
49
50 Still a preview — training continues; the two areas still being improved are
51 **audio** and **behaviour under fast, intense motion**.
52
53 ## Steps and strength — read this
54
55 - **4 steps is the recommended *minimum*; 4–8 is the useful range.** 6–8 steps
56 look noticeably better than 4, so add steps if you can afford them. Past **8
57 steps** it stops helping and can start to introduce **over-sharp artifacts** —
58 there's no benefit to going higher, so stay in **4–8**.
59 - **Keep strength at `1.0`.** It's tuned for 1.0 and holds up well across the 4–8
60 step range. Only reach for the strength dial if a *specific* clip misbehaves —
61 then **blurry ghosting / smear → nudge up** (`~1.05–1.2`), **over-sharp grain →
62 nudge down** (`~0.8–0.95`).
63 - Keep the scheduler on `simple`.
64
65 ## Use it in ComfyUI (recommended)
66
67 Custom nodes: **[Larryvrh/ComfyUI-MiniMax-H3-Turbo](https://github.com/Larryvrh/ComfyUI-MiniMax-H3-Turbo)**
68 — or search **"MiniMax-H3 Turbo"** in ComfyUI-Manager. (Keep the node updated; it
69 evolves alongside these weights.)
70
71 1. Install the nodes (Manager, or `git clone` into `ComfyUI/custom_nodes`) and put
72 a `.safetensors` from this repo into `ComfyUI/models/loras/`. You also need the
73 base MiniMax-H3 model, VAEs and text encoder — see the
74 [MiniMax-H3 tutorial](https://docs.comfy.org/tutorials/video/minimax/minimax-h3).
75 2. Start from the official MiniMax-H3 workflow (t2v or i2v) and make two changes:
76 - insert **MiniMax-H3 Turbo LoRA** between the model loader and the sampler;
77 - feed `SamplerCustomAdvanced` from **MiniMax-H3 Turbo Sampler**, and set the
78 scheduler to `simple` at **≥ 4 steps**.
79
80 Everything else stays as in the official graph, so both text-to-video and
81 image-to-video work. A ready-made t2v workflow ships in the
82 [node repo](https://github.com/Larryvrh/ComfyUI-MiniMax-H3-Turbo/tree/main/example_workflows)
83 (and here as `minimax_h3_t2v_turbo.json`) — drag it in.
84
85 - **Base model**: any MiniMax-H3 base — full (`bf16`, `int8_convrot`) **and the
86 pruned/curve variants** (`pruned_int8`, `pruned_fp8`). The node auto-detects a
87 pruned base and re-injects the time-conditioning at run time, so **one LoRA file
88 covers every base**.
89 - **`low_vram`** switch: **off** applies the LoRA at run time (sharpest,
90 recommended); **on** merges it into the weights for the lowest peak VRAM (a bit
91 softer on quantized bases). Turn it on only if you run out of memory.
92 - The custom sampler **auto-adapts to your ComfyUI version**: MiniMax-H3 runs
93 video and audio on two different flow schedules; recent ComfyUI handles that
94 natively (`ModelSamplingAV`) and older ComfyUI doesn't — the Turbo Sampler
95 detects which and does the right thing either way, so nothing to change when you
96 update ComfyUI.
97
98 ## Weights
99
100 All bf16, ~744 MB, applied as a plain low-rank update
101 (`W_eff = W + lora_B @ lora_A`, alpha = rank, so no extra scaling). **Prefer the
102 EMA files**; the non-EMA ones are for comparison.
103
104 | file | notes |
105 |---|---|
106 | **`minimax_h3_turbo_v4_step600_ema.safetensors`** | **recommended — current best.** Strong static/small-motion, good micro-detail, no over-sharpening. |
107 | `minimax_h3_turbo_v4_step600.safetensors` | v4-600 non-EMA (comparison). |
108 | `minimax_h3_turbo_v4_step150_ema.safetensors` | earlier v4 checkpoint. |
109 | `minimax_h3_turbo_4step_ema_ckpt850.safetensors` | `v1` line (~850) — over-sharpened / plastic in general, but the friendlier pick for **4-step heavy motion** (see above). |
110 | `minimax_h3_turbo_4step_ema_ckpt500.safetensors` | older `v1` (~500), softer. |
111 | `minimax_h3_turbo_4step_ema.safetensors` | initial release (~200). |
112
113 *Naming:* `v4` is the current training recipe and `stepN` is the training step.
114 Older files carry the previous `4step_ckptN` naming, where `4step` referred to the
115 sampler-step count.
116
117 ## Standalone (no ComfyUI graph)
118
119 `generate.py` is a single self-contained file — it loads the base DiT + a LoRA,
120 encodes the prompt, runs the few-step dual-schedule sampler, decodes and muxes an
121 mp4. It still needs a ComfyUI checkout for the H3 model / VAE / text-encoder
122 definitions:
123
124 ```bash
125 git clone https://github.com/comfyanonymous/ComfyUI
126 cd ComfyUI && pip install -r requirements.txt && cd ..
127 pip install -r requirements.txt # this repo: torch, safetensors, imageio-ffmpeg
128
129 # base weights from Comfy-Org/MiniMax-H3 into a models/ tree, then:
130 python generate.py \
131 --comfyui ./ComfyUI \
132 --base models/diffusion_models/minimax_h3_fl2va_bf16.safetensors \
133 --lora minimax_h3_turbo_v4_step600_ema.safetensors \
134 --te models/text_encoders/qwen3vl_32b_minimax_h3_int8_convrot.safetensors \
135 --video-vae models/vae/minimax_h3_video_vae_fp16.safetensors \
136 --audio-vae models/vae/minimax_h3_audio_vae_fp32.safetensors \
137 --prompt "A corgi in a chef hat flipping a pancake, sizzling sounds and a cheerful bark." \
138 --width 1344 --height 768 --frames 124 --steps 6 --out corgi.mp4
139 ```
140
141 ## Notes
142
143 - **Resolution / duration**: width and height are multiples of 32 (short edge
144 typically 768). Frame count is at 24 fps and snaps to the model's 17·k+5 grid
145 (124 ≈ 5 s). Validated range ~124–362 frames (~5–15 s).
146 - **VRAM**: the base model is large (~33 B); an 80 GB GPU is comfortable at the
147 largest resolutions. The ComfyUI node streams the base and adds the `low_vram`
148 switch, so it runs on much smaller GPUs. In the standalone script,
149 `--offload-adaln` trades ~13 GB of VRAM for CPU RAM.
150 - **Audio**: 32 kHz stereo, aligned to the video; the two streams ride different
151 flow schedules and are integrated each on its own clock. (Audio is one of the
152 two areas still being improved — see the top.)
153