FLUX 3 Max: Seedance-level image-to-video scores in about 12 seconds, now with 10-keyframe control in ComfyUI

FLUX 3 Max launched on Oct 19. fal Research post-trained it on top of Black Forest Labs' FLUX 3 base model. fal's launch post passed 1,400 likes and BFL's own post passed 980. On Oct 21 ComfyUI added it as Partner Nodes with two ready-made templates. Artificial Analysis lists it #2 on its AA-Video-I2V v1.1 board (statistically tied for #1) and #4 for text-to-video. All of this is preliminary, and the speed numbers below come from fal's own tests.

The headline is speed. fal's table has FLUX 3 Max and Seedance 2.5 both at 1,139 Elo for image-to-video. A 10-second 720p clip takes 12.45 seconds on FLUX 3 Max and 277.80 seconds on Seedance 2.5. (fal's press release gives 10.95 seconds, and its page gives 12.45, so call it 'about 12'.) MiniMax H3 sits just above at 1,141 but takes 83 seconds. In practice you can try twenty takes of a shot in the time one Seedance render comes back. fal lists $0.0425 per second at 720p until Nov 7, which is 75% off a $0.17 list price.

The ComfyUI nodes are where it gets useful for directors. The image-to-video node takes 1 to 10 keyframe images. With 'spread across the clip', the first image opens the shot, the last closes it and the rest are spaced evenly. With 'at times', you pin each image to a second you type, like 0, 2.5, 5. Clips run 5 to 20 seconds at 24 fps, in aspect ratios from 21:9 to 9:16. Synchronized audio (ambience, speech, effects) is on by default. Comfy's docs list 720p and 1080p, while fal's page describes 720p, so check what your endpoint actually returns. Note that these are cloud nodes that bill credits per second, not a local download.

Prompt habits the docs and fal's samples suggest: describe the sound in the same paragraph as the action, with ambience, music and speech kept separate so the layers don't fight. Write cuts explicitly as 'SHOT ONE … HARD CUT … SHOT TWO'. Leave reasoning_level on high for multi-subject or multi-shot prompts, which costs the same as low. fal's showcase prompts also block the shot out second by second ('seconds zero to two… seconds four to five…'). Limits: no continuation input (longer pieces are separate generations), audio can't be edited separately, and the seed doesn't make runs reproducible. Spotted, not tested. Credit fal Research and Black Forest Labs.