FL2VA / I2VA: mid-clip blackout ("iris-out" with a spotlight on the character), seed- and resolution-dependent
Summary
With the I2V/FL2V checkpoint, the room sometimes goes almost completely black around 0.5–1 s into a 5 s clip, only the character stays lit (like a spotlight / iris-out), and the scene returns when the end frame is reached. It depends strongly on the seed and the resolution (not on the prompt, LoRA, VAE, sampler, or model variant). At ~0.5 MP it never happened in my tests; at 1024x1280 it happened for 4 of 12 seeds. A milder version (background dims by ~10–25 luminance levels in the first ~1.5 s and stays dim) happens with I2V at 768x960 and above.
Setup
- ComfyUI 0.37.0, PyTorch 2.12 (cu128), RTX 5090, native MiniMaxH3 nodes (
MiniMaxH3ImageToVideo,BasicGuider,SamplerCustomAdvanced,res_multistep+simple, shift 12) - Model:
DasiwaMinimaxH3_dasiwaHybridV4_int8(also reproduced with the officialminimax_h3_fl2va_pruned_int8_convrotand DaSiWa V2 int8) - 4-step turbo LoRA (
minimax_h3_fl2v_turbo_4step_v1.2), audio off, 24 fps, 124 frames (5 s) - One anime-style test image (1024x1280, bright room with a window). FL2V: start = that image, end = a frame of the same scene with a different pose (colour-matched to the start)
Measurement (no frames needed): mean luminance of the frame's outer ring over time (64x36 grayscale). Plot attached: brightness_curves.png.
What I found (≈310 generations)
| Resolution | I2V, 5 s | FL2V (different end), 5 s |
|---|---|---|
| 640x800 | mild dimming (~-10) | fine (all seeds) |
| 768x960 | -23 avg (one seed: -48, ended dark) | fine (3 seeds) |
| 896x1120 | -18 avg | fine (3 seeds) |
| 1024x1280 | one seed: -138 / ended -51 | 4 of 12 seeds iris-out (outer ring drops to ~10/255) |
- FL2V at 1024x1280, 12 seeds: seeds 1001, 1111, 2222, 3333 fail; 2002, 3003, 4004, 5005, 6006, 7007, 8008, 9009 are fine.
- The failure starts ~0.5 s into a 5 s clip, ~1.75 s into an 8 s clip, and did not occur at 3 s (seed 1001).
- A failing seed fails with every setting I tried (below), including no turbo LoRA at all (20 steps, seed 1001: outer ring -171).
What did NOT help (all tried on failing seeds at 1024x1280; the iris-out stayed identical)
- Prompt: positive lighting wording, "continuous single shot", enumerating what must stay unchanged, timeline-style prompts,
live_wallpaper:form, flat-lighting wording, minimal prompt (9 variants × 4 seeds: 1/4 success at best, within noise) - LoRAs: Motion Booster V2 (0.5, 1.0); turbo LoRA strength 0.8/1.2; 8-step turbo; no turbo (20 steps)
- VAE: fp16 vs
int8_convrot(decoded results matched to 0.1 luminance — the darkening is already in the latent) - Models: official fl2va (pruned int8), DaSiWa V2 int8, DaSiWa V4
- Sampler/shift/attention: euler, shift 6, Sage Attention, TeaCache; adding a cross-faded keyframe at the midpoint
What helped
- Lower resolution (≤~0.5 MP) — stable in all my tests.
eulerinstead ofres_multisteproughly halves the mild I2V dimming (-10 → -5 at 640x800).dpmpp_2manduni_pcare worse;karrasremoves the dimming but visibly breaks the image.- Generating at low resolution and then re-sampling the upscaled video with the same model at denoise ≈0.3–0.45 (encode the frames into the video stream of the AV latent, as in
ComfyUI-H3-FaceRefine's inject node) gives a high-resolution result without the blackout (same seed that fails when generated directly).
Questions
- Is this a known property of the FL2VA/I2VA weights beyond the resolution the turbo LoRAs were trained for ("768p"), or is it something in the ComfyUI implementation? (I noticed #15426 about periodic dark frames every 17 frames; this looks like a different, seed-dependent failure.)
- Does the reference pipeline show it too at ~1.3 MP? Is there a recommended maximum resolution / token count for FL2VA?
- rkfg's ToneCompensate README mentions a systematic tone bias (denoiser runs darker than the anchored keyframes) — could this be the same root cause at a larger magnitude?
