ainvfx-fluid · synthetic training set

The training set of AInVFX/ainvfx-fluid, an IC-LoRA for LTX 2.5. An IC-LoRA is a small add-on model that makes LTX follow a control video. This one turns flat colour paintings into fire, smoke, explosions, steam, liquids, dust, fog and clouds. Every image and every video here is AI-generated. No camera footage and no stock clip was used.

Content

176 scenes. Each scene is one generated clip, shown to the trainer at three sizes, with nested frame ranges that all start at frame 0:

Folder	Training stage	16:9	9:16	1:1	Frames
high/	3 · fine detail	1920 x 1088	1088 x 1920	1408 x 1408	33
mid/	2 · shape and material	1024 x 576	576 x 1024	768 x 768	121
low/	1 · motion	576 x 320	320 x 576	448 x 448	321

Each stage folder has one subfolder per frame rate (24, 25 and 50). Each of those holds targets/<id>_<stage>.mp4 (the result) and control/<id>_<stage>.mp4 (the control video).

The rest of the repository:

• stills/: the first frames.
• manifest/<id>.json: every setting of one scene.
• manifest.jsonl: one line per training pair.
• scenes.json: the 176 scenes, with their prompts and seeds.
• keyframe_plan.json: the keyframes of every control.
• control_overrides.json: the doodle.py settings of the scenes whose controls were redrawn after review.
• low.toml, mid.toml, high.toml: the musubi-tuner dataset configs of the training run. Their paths point to /workspace/dataset/fluid_v3/, the folder they ran from.
• tools/: the scripts that made the set. They wrote to a staging folder, v3/, which this release moved to the root.

The mix:

• Fluids: 33 fire, 39 smoke, 28 explosions (3 of them nuclear, starting from their first instant), 30 steam, 20 liquids and water, 26 dust, fog, haze and clouds.
• Kind and scale: 56 elements on pure black and 120 full scenes, at small, medium and large scale.
• Format: 131 landscape, 31 portrait, 14 square.
• Frame rate: 71 at 24 fps, 68 at 25 fps, 37 at 50 fps.

How it was made

1. First frame (tools/stillgen.py): MOSS-Image SFT bf16 with the Qwen3.5 9B bf16 text encoder and the FLUX.2 small VAE. 40 steps, CFG 3, er_sde / simple, at 1920 x 1088, 1088 x 1920 or 1408 x 1408.

2. Video (tools/videogen.py): LTX 2.5 22B distilled bf16 with Lightricks' DFR recipe (Diffusion Fidelity Rendering), in ComfyUI v0.43.0 core nodes. Sampler euler_ancestral, CFG 1, first-frame strength 1.0.

   • Stage 1: half size, 321 frames, 8 steps, with one generated keyframe per second.
   • Stage 2: full size, on the first 121 frames. Latent x2, then the Pixel Spatial x2 IC-LoRA at 0.5, with the stage-1 latent as its reference. 3 steps, sigmas 0.909375, 0.725, 0.421875, 0.
   • 50 fps scenes: the motion is generated at 25 fps (161 frames). LTX's temporal x2 step then gives 321 frames at 50 fps. That step runs 4 steps, keeps the 25 fps keyframes pinned at strength 0.95, and conditions at 60 fps, as in Lightricks' reference pipeline. A direct 50 fps pass gave motion twice too fast.

3. Training clips:

   • high: frames 0 to 32 of stage 2.
   • mid: the 121 frames of stage 2, at mid size.
   • low: the 321 frames of stage 1 (or of the temporal step), at low size.

   Resizing uses lanczos, scaling to cover the frame then cropping at the centre. Encoding is H.264 at CRF 10.

4. Control (tools/doodle.py --keyframe-plan keyframe_plan.json): every keyframe is a single frame painted in flat colours, with black frames everywhere else. Keyframes sit on the grid of LTX's video VAE: frame 0, a group start (1, 9, 17 and so on) or the last frame, one per latent group, so at most 5 in a 33-frame clip and at most 8 elsewhere.

   • The split (keyframe_plan.json, written for the whole set at once): half the clips have a single keyframe, and half have several.
   • A single keyframe is the first frame, the last frame or a group start, a third of the time each.
   • Several keyframes follow one of three patterns, a third of the time each: the first and the last frame; the first and the last frame plus group starts in between (3 to 8 in all); or 2 to 8 keyframes anywhere.
   • Settings: full scenes use --mode scene --colors 6 --detail 6-10, elements on black --mode fluid --colors 4 --detail 4-10, and both --seed 1.
   • Redrawn after review: 41 scenes have their own settings, for a control closer to the result: thinner flames and drops, holes kept, faint smoke and steam kept, more colours for the tones of the plate. control_overrides.json lists them. They use doodle.py options that leave every other control unchanged: --detail up to 14, --keep-holes, --threshold, --opacity and --min-share for elements on black, and --scene-min-bin and --scene-min-share for full scenes.
   • Moved after a check: ST05_low had its single keyframe on frame 313, after the steam had gone, so the keyframe held no paint. It is now on frame 233. SM05 (faint smoke) left under 1 % of the frame painted in fluid mode; its three controls are now drawn in scene mode (--mode scene --colors 3 --detail 8).
   • Records: each manifest records the pattern, the keyframe frames, the options and the detail drawn for each clip.

5. Selection: 240 candidate scenes were generated. After review, 22 were dropped on their first frame, then 32 and 13 on their clips, and 3 nuclear scenes were added. 176 remain.

6. Captions: Qwen3-VL-8B-Instruct (the train caption command of ainvfx-runpod 1.2.0): about two sentences per clip on what the fluid is, how it moves, the setting, the light and the camera. The trigger ainvfxfluid, comes first. Each caption is one .txt file next to its target; its text is also in manifest.jsonl (caption_text) and in its scene's manifest.

Compute. About 14 GPU-hours on NVIDIA RTX PRO 6000 (Runpod), counting the first frames, the clips, setup time and the 50 fps redo. At the GPU's 600 W maximum board power, that is at most about 8.5 kWh; we did not meter it. The controls and the checks ran on CPU.

Training

The model was trained in three stages on these pairs only, each stage starting from the last checkpoint of the one before. The trainer is musubi-tuner (AkaneTendo25's fork, commit 9dc9abc) through the ltx_iclora recipe of ainvfx-runpod 1.2.0: LTX 2.5 22B dev, rank 32, alpha 32, the control at half size.

Stage	Clips	Steps	Learning rate
1 · motion	low/, 576 x 320 x 321	4,000	1e-4
2 · shape and material	mid/, 1024 x 576 x 121	2,000	5e-5
3 · fine detail	high/, 1920 x 1088 x 33	1,000	2e-5

The model card gives the full recipe, the time and the energy.

Licenses and provenance

• This set: CC BY 4.0.
• The first frames come from MOSS-Image by OpenMOSS (Apache 2.0).
• The videos come from LTX 2.5 by Lightricks, under the LTX-2.x Community License. That license claims no rights in outputs. A model trained on them is a derivative under its section 3.2.
• Every MP4 and PNG carries one AI-generation comment and nothing else. The prompts, seeds and settings are in the manifests.
• No real person, place or event is depicted on purpose.

This section is a summary, not legal advice. It follows the AInVFX statement of AI ethics.