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 smoke, steam, fire, explosions, clouds and fog. Every image and every video in this dataset is AI-generated. No camera footage and no stock clip was used.

Two kinds of shot are mixed on purpose, so the LoRA supports both uses:

• Elements on pure black, in the style of stock VFX elements, for compositing. A control painted on a black canvas should give an element on black.
• Fluids inside full scenes (sky, ground, landscape, city). A control painted with flat zones for the sky and the ground should give the whole shot.

108 training pairs are elements on black and 151 are full scenes.

Layout

Folder	Training stage	Size	Frames	Pairs (24 / 25 / 50 fps)
low/	1 · motion	576 x 320 (portrait 320 x 576)	321	89 (39 / 34 / 16)
mid/	2 · shape and material	1024 x 576 (portrait 576 x 1024)	121	86 (36 / 37 / 13)
high/	3 · fine detail	1920 x 1088 (portrait 1088 x 1920)	33	84 (32 / 40 / 12)
holdout/	samples only, never trained on	1024 x 576	121	4

Each folder holds one subfolder per frame rate. A pair is three files with the same name: targets/<id>.mp4 (the result), targets/<id>.txt (its caption) and control/<id>.mp4 (the control video). stills/ holds the first frame of every scene used here. manifest.jsonl has one line per clip: prompts, seeds, the generation and control settings, frame rate, size and the automatic QC numbers; its kind is black for an element on black and plate for a full scene. low.toml, mid.toml, high.toml and sample_prompts.txt are the musubi-tuner configs of the training run (their paths are those of the training pod; sample 5 uses the blank-canvas control of the ainvfx-fluid model card). tools/ holds the generation script, gen.py, the list of the scenes used here, scenes.json (prompts, seeds, frame rates), and doodle.py, the script that painted the control videos. gen.py makes the stills, the videos and their QC numbers; it uploaded each pod's results to a temporary generation branch, deleted after the review. The captions (step 4 below) were made by another script, not included here.

Every file on this branch has a role. The training run read every pair in low/, mid/ and high/. Its sample previews used the four holdout controls; the holdout results show what the generator made from the same stills, for comparison. stills/, manifest.jsonl and tools/ document how the pairs were made.

How it was made

1. Still. MOSS-Image SFT (the int8 file of Comfy-Org's repackage), 1920 x 1088 or 1088 x 1920, 40 steps, CFG 3, sampler er_sde, scheduler simple. 101 scenes generated in two rounds, 95 of them used here. Round 1: 28 elements on black, 32 full scenes and 3 holdout scenes. Round 2: 38 scenes, mostly large billowing smoke (thick, soft, rounded masses with diffuse edges) plus the gaps of round 1: explosions in full scenes, more kinds of fire, dust, sea fog and clouds seen from above.
2. Video. LTX 2.5 22B distilled (int8), image to video from the still, resized to each stage's size: 8 steps, CFG 1, euler_ancestral, first-frame strength 0.7, LTXV Preprocess at 18. The 1920 x 1088 clips use two passes (960 x 544, then the x2 latent upscaler and 3 refine steps), as Comfy-Org's video_ltx2_5_i2v template does. ComfyUI v0.43.0, core nodes only.
3. Control video. Each clip is painted over automatically in flat colours, the way an artist would block it in. Full scenes: 6 flat colours chosen among the colours of the frame, every zone filled (sky, ground, smoke, fire). Elements on black: black plus up to 3 flat colours, one blob per volume. Each control has 2 to 5 keyframes on the 8-frame grid of the LTX video VAE (the first and the last frame always), each held for 1 to 8 frames, and black frames everywhere else. The script that paints them is tools/doodle.py. Full scenes use --mode scene --colors 6 --detail 6-10 --keyframes 5 --hold 1-8 --seed 1, elements on black --mode fluid --colors 4 --detail 4-10 --keyframes 5 --hold 1-8 --seed 1. Detail and hold are ranges: each clip draws its own values from a seed made of --seed and its file name, and control_settings in its manifest line records them.
4. Caption. Qwen3-VL-8B-Instruct describes each clip in about two sentences: what the fluid is, how it moves, the setting, the light and the camera. The trigger ainvfxfluid, comes first. Words about style or medium (photorealistic, cinematic, CGI, footage, rendered and similar) are removed, so the trigger carries the look and the caption carries the content.
5. Selection. Clips with no motion, a cut, a drifting background or a visible artefact were removed after review. After the first training run, large-scale smoke with thin wisps and fine curls (it read as cigarette smoke) was removed too: B01, B02, B07, B25, B26 (B01 stays only as a holdout sample). Small wisps inside a larger shot are kept.

Compute. Generation ran on five NVIDIA RTX PRO 6000 (96 GB) on Runpod in parallel, 15 to 28 minutes each from pod creation to verified upload (setup and model downloads included), over two rounds of five pods, 3.5 GPU-hours in all. At the GPU's 600 W maximum board power, that is at most about 2.1 kWh; we did not meter it.

Licenses and provenance

• This dataset: CC BY 4.0.
• The stills come from MOSS-Image by OpenMOSS, released under Apache 2.0 (weights and code).
• The videos come from LTX 2.5 by Lightricks, under the LTX-2.x Community License. That license claims no rights in outputs (section 5). A model trained on these videos is a derivative of LTX-2.x under its section 3.2, as the IC-LoRA already is.
• The captions come from Qwen3-VL-8B-Instruct, Apache 2.0.
• Every MP4 and PNG carries one metadata comment: an AI-generation label that names the models. The ComfyUI generation graphs were removed from the files; the prompts, seeds and settings are in manifest.jsonl.
• No real person, place or event is depicted on purpose. This card is the AI-generation disclosure for every file.

This section is a summary, not legal advice. It follows the AInVFX statement of AI ethics: documented training data, labelled generated content, stated compute.

Who made it

Made by AInVFX. We teach generative AI for film and TV in the Generative AI Bootcamp for Film and TV: 60 hours of live training over several weeks, online and in English, built around your own project, with cloud GPU included. It runs several times a year: see the upcoming cohorts on ainvfx.com/bootcamp.

Questions, results, bugs: contact AInVFX or open a discussion on this repo.