VFX 4 All

Cinematic VFX presets for everyone. Drop a photo or a video clip, pick an effect, and get a film-grade VFX shot with sound: a walk through fire, a body turning to liquid chrome, a clone army, a city folding overhead.

VFX 4 All is an open source app built on the Comfy Developer Platform. It runs open-weights models (LTX-2.5 and Qwen Image Edit 2511) through the official Comfy SDK, either on Comfy Cloud or on your own ComfyUI.

Walk of Fire	Liquid Chrome	Synchronized	The Fold
Disintegrate	Colossus	Zero Gravity	Frozen Rain (beta)

Try it

• In your browser, no key: vfx.nomastudio.ai plays every effect. Pick one and open it as a Comfy Cloud app: sign in with your Comfy account, add a photo and run it. The run bills your own Comfy account.
• On your machine, everything: the full app adds video mode, a second photo for the likeness, 9:16 and HQ. It runs with your own Comfy API key, see Run it.

Two ways in

From a photo. Upload one or two photos of a person. Each preset builds its own film scene around them (location, wardrobe, lens, light), keeps their face, and then animates the effect. You can also skip the scene and animate your own image as it is.

From a video. Upload a clip you filmed or generated, scrub to the moment the effect should start, and VFX 4 All continues the shot from that exact frame. The result is your clip up to the cut, then the effect, joined on the same frame so the cut is invisible. Clips from 2 to 60 seconds, 480p to 4K, up to 500 MB (VIDEO_LIMITS in vfx4all/config.py).

No photo or clip at hand? Each mode loads our own samples with one click (see samples/). Under More options, an extra direction in English goes after the preset's effect prompt, and a seed repeats a take.

In Comfy Cloud, no key at all. Every preset that animates a single first frame (41 of the 45) can also run as a Comfy Cloud App Mode app: one graph that takes your photo, builds the scene and renders the effect. You open it with your own Comfy account and run it with your own credits; nobody hands an API key to anyone. tools/build_app_graph.py builds those graphs from the presets, and each preset keeps the share link of its app in comfy_app.

How it works

photo(s) --> Qwen Image Edit 2511 --> first frame of the preset's scene --+
                                                                         +--> LTX-2.5 DFR image to video --> shot with audio
your image (as it is) ---------------------------------------------------+
                                                                         |
video clip --> frame at the chosen second -------------------------------+--> splice back into the clip

first frame --> Qwen Image Edit 2511 (or a mirror) --> last frame --> LTX-2.5 first and last frame --> shot with audio

• workflows/ holds the ComfyUI graphs in API format: the scene, the LTX-2.5 DFR effect (default) and the classic two pass effect (VFX4ALL_EFFECT_PIPELINE=classic). See workflows/README.md for every node the app sets.
• presets/ holds one JSON file per effect: a scene prompt, an effect prompt written for that scene, and a generic prompt for user images and video splices. Copy one to make your own. A preset can also ask for a last frame (last_frame_prompt, edited from the first frame with Qwen Image Edit 2511, or last_frame_mode: "mirror") to be animated first-to-last, a LoRA from the Comfy Cloud catalog (effect_lora), or to be played backward after rendering (post: "reverse").
• vfx4all/ is a small FastAPI server. It submits the graphs with comfy-sdk, follows the jobs (and cancels them), checks every take (a black or NaN take is retried with the next seed), repeats the uploads and downloads that time out, and splices videos with ffmpeg.
• web/ is the interface: plain HTML, CSS and JavaScript, no build step. It shows the prompt that was sent, keeps your results across reloads and compares a spliced clip with your original.
• tools/build_app_graph.py builds the App Mode graphs, tools/build_site.py builds the static catalog site and tools/batch_presets.py renders test versions of the presets; tools/dev/ keeps the A/B scripts behind the choices in workflows/README.md.

Run it

Python 3.10 or newer.

git clone <this repo> vfx4all
cd vfx4all
python -m venv .venv
.venv/Scripts/pip install -r requirements.txt   # on macOS or Linux: .venv/bin/pip
cp .env.example .env                            # then paste your COMFY_API_KEY
.venv/Scripts/python run.py                     # opens http://127.0.0.1:8420

ffmpeg ships inside the imageio-ffmpeg wheel, so there is nothing else to install. requirements.txt pins the exact versions the app is tested with.

No photo or clip at hand? samples/ has two photos and two clips to try both modes; see samples/README.md for which effect to pair with each.

Run the tests

.venv/Scripts/python -m unittest -v

The tests run offline in a few seconds. They check every preset and graph, the App Mode builder, the static site build, the upload limits, the ffmpeg steps (phone clips shot vertical included), the HTTP API with its hosted mode, access code and daily cap, the job queue, and that every model the graphs load is listed in THIRD_PARTY_NOTICES.md. They never reach Comfy Cloud (the test package blanks COMFY_API_KEY before the app loads), so they cost nothing. GitHub Actions runs them on Linux and Windows with Python 3.10 and 3.12. Only a real render tells you whether a prompt works.

Who pays for the generations

Every generation bills the Comfy Cloud workspace that owns the API key it runs with.

• Running it yourself: you put your own key in .env, so you pay for your own runs. API access needs a paid Comfy Cloud plan. Keys come from platform.comfy.org.
• Sharing it without keys: link each preset to its Comfy Cloud App Mode app (the comfy_app field of the preset). Visitors sign in to Comfy and their runs bill their own account.
• Hosting it for others: visitors can paste their own key in Advanced. The key stays in their browser and travels in the X-Comfy-Key header, so their runs bill their workspace; the server uses it for that job only and never stores it. If they leave it empty, the server key from .env is used, so read Hosting it before you expose the app.
• Your own GPU, no cloud credits: the models are open weights, so the same app can drive your own ComfyUI through comfy-api-proxy, with no key. Set COMFY_BASE_URL=http://127.0.0.1:8189 in .env.
• A Comfy deployment: a ComfyUI of your own on a dedicated cloud GPU, billed while it is up. Set COMFY_BASE_URL to its URL.

docs/RUN_ON_YOUR_COMFYUI.md lists every model file the graphs load, with its source, revision and download command, and the steps for the proxy and for a deployment. The top bar of the app says where it runs.

Only open-weights models run in the workflows, so no partner API credits are used: Comfy Cloud charges GPU time. In our tests on Comfy Cloud a scene took about 10 seconds of GPU and a 6 second Draft effect between 12 and 30 seconds.

Presets

45 presets in 8 categories:

Category	Presets
Elements	Walk of Fire, Disintegrate, Ice Age, Lightning Strike, Smoke Exit, Water Body, Sandstorm, Ember Wings
Transformations	Liquid Chrome, Gold Touch, Stone Statue, Glass Shatter, Paper Fold, Inflate, Hologram, Wings of Light
Clones	Synchronized, Mirror Hall, Crowd of Me, Shadow Twin
Worlds	The Fold, Earth Zoom, Portal Jump, Mini World, Mirror City, Lidar Scan, Paint World
Scale	Colossus, Ant Size, Pocket Giant
Physics	Zero Gravity, Gravity Flip, Bullet Dodge, Super Jump, Shockwave
Time	Frozen Rain, Bullet Time, Rewind, Aging, Stop Motion
Style	Comic Panel, Anime Shift, Claymation, Neon Outline, Film Noir

Walk of Fire, Liquid Chrome, Synchronized, The Fold, Disintegrate, Colossus and Zero Gravity are stable; the other 38 are marked beta in the app (see Limits). Most presets take all three inputs. These run only from a photo, because the effect depends on its own scene or on a last frame built from it: Colossus, Frozen Rain, Mirror Hall, Shadow Twin, Ant Size, Pocket Giant, Bullet Time and Comic Panel. Rewind takes a photo or your own image, not a video.

Make your own preset

Add a JSON file to presets/:

{
  "id": "my-effect",
  "name": "My Effect",
  "category": "Elements",
  "description": "One line for the card.",
  "modes": ["scene", "own_image", "video"],
  "duration": 6,
  "wardrobe": {"man": "...", "woman": "...", "person": "..."},
  "scene_prompt": "Create a new cinematic film still in {aspect_words} format. {references} The {subject} is the only person in the new still, with exactly {their} face, eyes, nose, jaw, hairline, hairstyle, hair colour{facial_hair}, skin tone and build{identity_notes}. The clothing, the place and the light come only from this description. ...",
  "effect_prompt": "Use the provided start image as the first frame. ...",
  "generic_prompt": "Use the provided start image as the first frame. ..."
}

Placeholders: {subject} (man, woman or person), {their}, {Their}, {them}, {wardrobe}, {aspect_words}, {identity_notes}, {references} and {facial_hair}. {references} introduces the photos the way Qwen Image Edit 2511 reads them ("Image 1 and image 2 are two photos of the same man. Use them only as the identity reference.", or the one-photo version), and {facial_hair} adds ", facial hair" for a man only, so a woman's scene is never asked to keep a beard. Prompts avoid subject pronouns so the verbs agree for all three subjects. A thumbnail at web/thumbs/<id>.jpg and a short loop at web/previews/<id>.mp4 make the card.

Prompting notes that made these presets work: describe the state the shot starts in rather than the transition into it, give every effect a sound, lock counts and scale with a named rule (for example "CLONE COUNT LOCK: seven figures in the final frame"), and keep every instruction positive. For scenes, name each photo with its role and keep the identity block above as it is: with the old wording ("the person from the reference images") Qwen Image Edit 2511 exaggerated grey hair and earrings. In wide full-body shots the face gets small and likeness drops, so frame the subject closer when the effect allows it.

Quality

Draft renders at 896x512 (16:9) and HQ at 1280x704, both at 24 fps. The effect runs through LTX-2.5 DFR (Diffusion Fidelity Rendering): detailing keyframes plus a spatial x2 pass with the Refine-Details IC-LoRA. Against the classic graph it keeps faces and textures much sharper; the A/B is in workflows/README.md. For delivery, upscale the shots you keep, for example with the RTX Video nodes in ComfyUI.

Limits

What VFX 4 All does not do well yet, from our own test runs:

• Likeness drifts in the strongest transformations. The scene keeps the face from your photo, but effects that turn the whole body into another material or style (Stone Statue, Mirror Hall, Anime Shift) lose part of the likeness by the last frame. Start from a sharp photo where the face is clearly visible.
• One person per run. Presets are written for a single subject (man, woman or person). Group photos, pets and objects are not supported; the clone presets make the copies themselves.
• Short shots. An effect lasts 4 to 6 seconds (the preset's duration) at 24 fps. Draft is 896x512 and HQ 1280x704, or the vertical equivalents. Upscale the shots you keep in another tool.
• Video mode continues one frame. The effect is rendered from the frame at the cut, so it does not inherit the camera move or the motion of your clip, and its sound is generated. 36 of the 45 presets take a video. Clips must be 2 to 60 seconds, 480p to 4K and up to 500 MB.
• Black takes. Comfy Cloud sometimes returns an all black take or stops with a NaN error, and in our tests these came in waves. The app checks every take and retries with a new seed (for most presets two DFR attempts, then the classic graph); when every attempt fails, the job fails. Every attempt uses GPU time, so a run that needed retries costs more.
• Beta presets. 38 of the 45 presets are marked beta: they work but vary more between takes, and some show artifacts (a stray pillar in Earth Zoom, rough paper in Paper Fold). Try two or three seeds.
• Last frame presets take an extra step. Ant Size and Comic Panel first edit a last frame with Qwen Image Edit 2511 (Bullet Time mirrors the first one) and then animate first to last, so they only run from a photo scene and spend one more image generation. Rewind renders forward and plays the result backward, sound included.
• App Mode covers 41 presets. Ant Size, Bullet Time, Comic Panel and Rewind need more than one graph or a step after rendering, so they only run in the VFX 4 All server.
• Comfy Cloud model names. The graphs load models by their Comfy Cloud catalog names. On your own ComfyUI you need the same files under the same names, and a rename in the catalog breaks a graph until it is updated.
• Commercial use depends on the models. Photo scenes and last frames come from Qwen Image Edit 2511 (Apache 2.0). LTX-2.5 is free only for entities under USD 10,000,000 in annual revenue. The Hologram preset uses a style LoRA with no published license. Details in THIRD_PARTY_NOTICES.md.
• No cost display. The app does not estimate what a run costs. Check the usage of your Comfy Cloud workspace.
• One server, one process. Jobs live in memory and in their job.json, so run a single instance. Finished jobs survive a restart; a job the restart interrupted is marked failed with the ids of its Comfy Cloud jobs, whose results may still be in your Comfy history.

Hosting it

By default VFX 4 All is a local app: it listens on 127.0.0.1, has no login and keeps every run in runs/. To open it to other people, set VFX4ALL_MODE=hosted (see .env.example):

• Jobs, uploads and results belong to the browser session that made them; nobody else can list or open them, and job.json and error logs are never served.
• Clips are capped at 30 s and 100 MB, a session runs two jobs at a time (VFX4ALL_SESSION_JOBS) and one network six (VFX4ALL_IP_JOBS), and runs and uploads older than 48 hours are deleted (VFX4ALL_RETENTION_HOURS).
• VFX4ALL_ACCESS_CODE asks for a code before any upload or generation, for example a demo for a jury. Visitors enter it in Advanced.
• VFX4ALL_DAILY_RUNS caps the generations the server accepts per UTC day, so a shared link can never spend more than you planned. Set it whenever the server key pays.

The Dockerfile builds that setup (hosted mode, port 8420, runs on a /data/runs volume). Put it behind a reverse proxy with HTTPS and set FORWARDED_ALLOW_IPS to the proxy's address. The server sends a strict Content Security Policy and serves the fonts itself, so the page loads nothing from other sites. If you host it for others, the LTX license asks you to put its Section 4 and the Acceptable Use Policy into your own terms (see THIRD_PARTY_NOTICES.md).

The catalog site

A public page that spends nothing does not need the server at all. tools/build_site.py writes site/: the same interface in static mode, where every card plays its preview and opens the preset's Comfy Cloud app (comfy_app), so each visitor runs it with their own Comfy account. The studio and the key settings are left out, presets.json replaces the API, and vercel.json sends the same security headers as the server.

.venv/Scripts/python tools/build_site.py --demo path/to/demo.mp4   # --demo is optional
cd site && vercel deploy                                           # or any static host

On Vercel, the first deployment of a project goes to production and takes the project's domains, even without --prod.

Consent

Only use photos and videos of people who agreed to it. The app asks you to confirm that before every upload, and sends what you upload to the Comfy workspace of the key in use.

License

MIT for the code in this repository. The models, ffmpeg, the NoMa Studio AI brand assets and the demo media keep their own licenses or owners: see THIRD_PARTY_NOTICES.md, and read the model licenses before you use the results commercially.

Made by NoMa Studio AI for the #OpenSourceHiggsfield challenge.