KennethFal/vh5tape-vhs-lora-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

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

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. Stay near 1 MP (very low resolutions are unstable) and keep prompts to a single shot.

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.