Paint a plate in action: a showcase, new features and an open dataset for ainvfx-fluid

A week ago we submitted Paint a plate to Comfy's Open Source Higgsfield contest. Paint a plate is our open-source app for VFX artists, built on top of ComfyUI: you paint on the first frame of a plate, and the edit carries through the whole shot. The code is now frozen until the contest review ends, so we spent the last few days using the app instead of improving it further. Here are four minutes of what it makes, what changed in it before the freeze, and the new version of ainvfx-fluid, retrained only on synthetic data that we publish.

🎬 Watch the showcase

Every clip has been generated live with the app. In order:

  • A candle, painted on a blank plate, becomes a real one. The Color Inspector then reads its flame HDR values, above the 1.0 white of a standard video=.
  • The training set of ainvfx-fluid: training pairs side by side, then pages of the generated clips it learned from.
  • A nuclear blast, painted the same way, then exported as EXR sequences, a combined clip and a licence manifest.
  • Four fluids at once: a wave and a waterfall, red and yellow smoke over water, an explosion with fire trails, a plume of smoke.
  • A fire monster, painted on two keyframes.
  • A camera rig erased, then the set extended into a full corridor.
  • A green screen replaced by a ruined city, then the car set on fire.
  • A lamp added, the scene relit for the evening, then each layered edit switched off and on again.
  • A rough boat that comes back in the plate's own 2D style, beside a robotic arm that follows the gardener's motion.

All the source plates are synthetic, from our test dataset.

🆕 What changed since our initial release

Paint a plate went from version 1.0 on 30 October to 1.11 early this week. The changes that matter most:

  • Paint in the image's own style. Paint follows your drawing's outline, pose and colours, in the style of the image: a photograph, a painting or an illustration. The 2D boat in the showcase comes from a rough sketch.
  • Relight from a light plan. Pick a colour and paint where the light comes from. The source itself appears there (a lamp, a window, a fire, the sun), and the light falls on what faces it. Paint in black to darken an area.
  • A player that reads like OpenRV. F5 opens the Color Inspector, which reads a pixel's own values, as in RV. Exposure, contrast, gamma, the channels and the zoom work as RV's.
  • Find where a fault comes from. The Render panel lists every pass of a render. Ctrl+click puts one pass against another in the wipe, and W, the X-ray, tints where the edits change the plate.
  • HDR where the plate is HDR. On a plate that holds values above 1.0, such as an EXR sequence or a log video, new content goes through Lightricks' SDR-to-HDR IC-LoRA, then sits at the plate's exposure.
  • Fluids in half float. The app now opens on a blank plate, and a fluid shot comes out as half-float EXR frames when the SDR-to-HDR IC-LoRA is installed.
  • A licence manifest in every export. Each export lists the models behind the result with their licences, in a MODELS.md file.

🔥 ainvfx-fluid, retrained on an open synthetic dataset

ainvfx-fluid is our free IC-LoRA for LTX 2.5: a small add-on model that turns flat colour blobs, painted on a few keyframes, into smoke, steam, fire, explosions, water and other liquids, dust or clouds. Paint a plate uses it for every fluid shot.

Please update it. An earlier version was trained by mistake on stock clips whose terms did not allow it. We withdrew it, and on 1 November we retrained ainvfx-fluid only on synthetic clips that we generated and publish. If you downloaded it before 1 November 2026, delete that file and download it again. Paint a plate uses the new file since version 1.0.5.

The training set is open, on Hugging Face under CC BY 4.0:

  • 176 scenes, 528 training pairs: 56 elements on black and 120 full scenes, from fire to liquids, at 24, 25 and 50 frames per second.
  • Every clip is synthetic. MOSS-Image made each first frame, LTX 2.5 made each clip, and a script painted each control, the flat-colour video that the model learns to follow.
  • Every clip is documented: its prompt, seed, models and settings. The scripts that made the set are in it too.
  • Every file is labelled as AI-generated. No stock footage, no scraped content, no real people.
  • The compute is stated: 16 h 20 min of training on one NVIDIA RTX PRO 6000, about 33 GPU-hours in all with the data, at most about 20 kWh.

This is what clear provenance means to us: anyone can check what the model learned from, and rebuild the set.

💻 Try it

Paint a plate needs ComfyUI v0.43.0 or newer, the ComfyUI-LTXVideo nodes and an NVIDIA GPU, and we recommend 32 GB of VRAM or more. Install it from ComfyUI Extension Manager (search for Paint a plate), by hand, or on Runpod. The repository and our first post give every step, and our 23 test plates are on Hugging Face.

🤝 Tell us what you make

Comfy announces the winners on 10 November. Until then, the code stays as it is and we keep painting. Show us your shots in the comments, or report what breaks in the repository's issues. If you share your shots, tag #OpenSourceHiggsfield.

Interested in learning more? Our education programs help you build workflows like these. Come build with us!

🔗 Sources & Links

🖌️ Paint a plate:

🔥 ainvfx-fluid:

🏆 The contest:

🧠 Models and tools:

🎵 Music:

💬 Community:

📄 Related Posts:

Join the conversation

Have thoughts on this article? We'd love to hear from you!

Let's work together

Inquire now