ComfyUI: CFG-distilled sampling recipe (BasicGuider, shift 12/3) + load-console and strength-0 diagnostics

d25a4ab verified 21 days ago

7.59 kB

license: other
license_name: minimax-h3-community-license
license_link: https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE
base_model: MiniMaxAI/MiniMax-H3
base_model_relation: adapter
pipeline_tag: image-text-to-video
tags:
  - lora
  - video
  - text-to-video
  - vhs
  - retro
  - 1980s
  - style
  - minimax-h3

vh5tape — a worn-VHS look for MiniMax H3

A style LoRA that makes H3 footage look like it was recorded off 1980s broadcast television onto a VHS tape that has seen better days — soft smeared detail, chroma bleed, tracking noise, head-switching bands at the frame edge, and (because H3 trains audio jointly) the matching muffled mono sound, tape hiss and warble.

The damage is a prompt-controlled dial with three trained levels, from a gentle analog glow to a tape the VCR can barely hold onto.

Gallery

All clips are raw single takes at 480P / 4:3 — none of this content was in the training data.

Prompting

Prompt the way the model was trained: trigger first, damage tier second, scene third.

vh5tape, badly damaged VHS tape with heavy tracking errors and distortion. A loud
1980s local used car commercial: a salesman in a plaid suit shouts at a car lot
with strings of pennant flags and a spinning price sign, cheesy synthesizer
jingle, echoing announcer voice.

The damage dial

level	phrase
light	`lightly worn VHS tape with slight analog noise`
medium	`worn VHS tape recording with visible tape damage`
heavy	`badly damaged VHS tape with heavy tracking errors and distortion`

Tips

Trigger: vh5tape — put it at the very start of the prompt.
Prompt expansion: the LoRA was trained through H3's caption format, and the balanced expansion mode preserves the trigger and damage phrases — safe to keep on.
Scale: 1.0 for the full look; 0.7–0.9 softens the wear if heavy prompts overshoot; up to ~1.5 pushes toward barely-watchable tape (fun for horror).
Sound comes along for free: describe the audio ("muffled mono audio", "canned laughter", "cheesy jingle") and it renders appropriately tape-worn.
Dialogue: give characters short explicit quoted lines (the host says: "It slices, it dices!") — they come through near-verbatim.
Period set dressing helps: wood paneling, studio lighting, 4:3, era wardrobe.

Recipe (fal)

Works on the H3 and H3-Max LoRA endpoints (minimax/h3/text-to-video/lora, minimax/h3-max/text-to-video/lora) and their image-to-video LoRA siblings:

{
  "prompt": "vh5tape, worn VHS tape recording with visible tape damage. A 1980s ...",
  "loras": [{ "path": "<url to vh5tape.safetensors>", "scale": 1.0 }],
  "resolution": "480P",
  "aspect_ratio": "4:3",
  "duration": 15,
  "prompt_expansion_mode": "balanced"
}

480P / 4:3 is the period-correct sweet spot; the style carries fine to higher resolutions if you want clean pixels of a dirty tape.

ComfyUI

vh5tape.safetensors loads in ComfyUI as-is — no conversion needed. It uses the ComfyUI-native key layout (diffusion_model.blocks.N.attn.{qkv_proj,out_proj}, fused QKV, alpha = rank), so all 104 modules load with no warnings.

Base checkpoint — this is the thing that matters. Use a non-pruned, non-rotated base:

✅ minimax_h3_fl2va_bf16.safetensors (recommended)
✅ minimax_h3_fl2va_pruned_fp8_scaled.safetensors
❌ *_int8_convrot, nvfp4, w4a8 — these store weights in a rotated / quantized basis. The LoRA loads without any error and then produces warped faces, melting limbs and disappearing objects. The default checkpoint in the official ComfyUI H3 tutorial is minimax_h3_fl2va_pruned_int8_convrot — you must change it. (Not specific to this LoRA — MiniMax-H3 LoRAs in general are made for fp16/fp8 bases.)

Setup

Put vh5tape.safetensors in ComfyUI/models/loras/.
UNETLoader (bf16 or fp8_scaled fl2va) → LoraLoaderModelOnly → MiniMaxH3ImageToVideo / your sampler.
Strength 1.0 (useful range 0.7–1.2).

Sampling recipe — H3 is CFG-distilled, so SDXL/Flux habits will degrade it with or without a LoRA: use BasicGuider (not CFGGuider), res_multistep + simple, ~20 steps, video shift 12 / audio shift 3 (the MiniMax H3 Sigma Shift node). Stay near 1 MP (very low resolutions are unstable) and keep prompts to a single shot.

Stacking with a turbo LoRA: drop this one to 0.4–0.6 — the two perturbations add, and 4-step turbo has little headroom. (The turbo_mode checkbox occupies its own LoRA slot; add a second LoraLoaderModelOnly for this one.) 4-step turbo also softens the tape grain — ~20–25 full-precision steps show the style best.

If it looks wrong, two 2-minute checks: (1) watch the console at load — this LoRA should patch with zero "lora key not loaded" lines; (2) on a quantized base, load the LoRA at strength 0.0 and compare to no-LoRA — if output degrades even at 0.0, the damage is the quantize→merge→requantize round-trip on your base checkpoint, not this LoRA: switch to the bf16 base.

vh5tape-comfyui.safetensors is an optional convenience build: bit-identical weights plus explicit .alpha tensors (= rank, scale 1.0) and the base-model requirement in its metadata — same look, just harder to mis-load in third-party loaders.

Training

Trained with fal's MiniMax H3 t2v trainer (rank 32, 5,000 steps, 4:3, joint video+audio objective) on a small hand-curated set of four-second clips of period broadcast material — sitcoms, soap operas, commercials, news and station sign-offs — each captioned with its content and one of the three damage-tier phrases above. The 5,000-step checkpoint beat its 2k/3k/4k/6k siblings in blind A/B testing; more steps past 5k made the damage tamer, not heavier.

Get the weights

file	what
`vh5tape.safetensors`	the LoRA, rank 32, 5,000 steps
`vh5tape-comfyui.safetensors`	same weights bit-exact + explicit alpha tensors, for ComfyUI (see the ComfyUI section)
`PROMPTS.md`	the exact prompts + seeds behind every sample above

Made for H3-TV — a 1970s portable CRT being rebuilt to play an endless, slightly wrong 1980s broadcast day.