One HTML file that turns your phone into a remote for ComfyUI. Queue generations, swap models and LoRAs, and watch images and video land in a live gallery while your own PC does the heavy lifting.
ComfyUI is the most capable open-source image and video engine there is, and it asks you to sit at your PC, staring at a node graph, for every single generation. Cloud services solved mobility by charging monthly rent for compute you already own.
The missing piece was never more GPU. It was reach: something with the immediacy of a Walkman. Personal, portable, instant. Existing remote tools dumped every node parameter into raw unsorted lists; technically universal, practically unusable on a phone.
A complete ComfyUI companion that runs in any mobile browser. No app store, no install, no server-side code. Drop a workflow PNG or JSON on it and every meaningful parameter is auto-detected, intelligently sorted, and styled for one-handed use. Anything unrecognized lands in a collapsed Advanced section instead of disappearing.
Fields sort into a consistent hierarchy: model, seed with a 🔒 lock for A/B testing, prompt, sampler, compact parameter rows. Checkpoints, UNETs, and GGUFs all populate their dropdowns, organized by folder.
Generations stream in with slide animations. Same-prompt runs stack into fanned-card albums. Videos auto-play as they cross the center of the screen and loop in the lightbox.
Tap ⌬ on any LoRA and its CivitAI page comes to you: trigger words as tappable glowing bubbles, example images with full generation metadata, one-tap prompt harvesting into your history.
Every output carries its workflow in embedded metadata. The ∞ button reloads it, seed preserved. Blended prompts wear the same mark in your history, glowing in your theme.
Wan 2.2 image-to-video with the Lightning setup ready in the example workflows. Frame interpolation for smoother motion is one toggle; the app rewires the graph on submit.
The app adds zero content filtering of its own. Every built-in prompt defers entirely to the model you chose to run on your own hardware. Your rig, your rules.
The Tools pane drives the LLM running in LM Studio on your own PC, through ComfyUI. No API keys, no cloud, no per-token bill.
Natural language in, structured JSON prompt out. Add up to four concepts and blend them into one scene.
Reverse-engineer prompts from reference images. Multiple images get analyzed individually, then fused.
Talk directly to your model, from your phone. Full conversation memory, image attachments, no middleman.
An 18-field character builder. Output instant JSON or LLM-composed prose; your roster saves to history.
The system borrows from a photographic darkroom and a holographic trading card at the same time, and keeps them strictly separated:
The theme accent. It marks everything live: progress shimmer, active states, glowing glyphs, the running job. Seven themes recolor it; this entire page obeys the same rule. Try the swatches.
Iridescent translucency. It marks structure: card surfaces, frames, the ∞ emblem. It never changes with the theme, like a holographic foil that reads the same under any light.
Developed through rapid sessions with Claude: describe a behavior, test it on a real phone against a live ComfyUI server, report what broke, refine. Over sixty archived builds document the evolution. Every design call, from animation timing to touch targets, was tuned against real fingers on real hardware.
WebSocket link to ComfyUI, PNG metadata extraction, and the auto-detection system that turns any workflow graph into an editable mobile form.
Mobile browsers kill background connections. Missed-generation sync, auto-reconnect, and persistent state made the gallery trustworthy: lock your phone mid-render, come back, everything's there.
JSON import with a full UI-to-API converter, resolution of bypassed and rerouted nodes, GGUF loaders, and an Advanced section so unknown custom nodes expose their parameters instead of vanishing.
Shimmer progress with live ETA, slide-in gallery animations, long-press auto-pan, seven glowing themes. Speed stayed sacred; every effect got tested for snappiness and cut if it slowed a tap.
Text and image prompt generation grew into a four-tool suite: multi-concept blending, a full chat client, a character creator, and CivitAI LoRA intelligence, all riding the model on your own machine.
Errors that name the failing node so a fix takes one report, example workflows for five model families forming an end-to-end demo chain, and docs written for people who already know their way around ComfyUI.
No subscription, no upload queue, no one else's servers. The most capable image engine you can run is the one you already own. Your GPU, in your pocket.
Released as free, open-source software for the ComfyUI community. One file you can read top to bottom. Development continues in public, with beta testers surfacing new node types and workflows to support. Supported by donations, not subscriptions.
