A private, self-hosted image and video generator for the Teczo team, similar in idea to Higgsfield. You pick a preset, type a prompt, and press Generate. ComfyUI does the work on our own GPU.
This file is an audit of the app as of 2026-10-07. It covers what is built, how it works, what is
weak or risky, and what is coming next. It is written for brainstorming improvements.
CLAUDE.md is still the source of truth for rules and decisions.
| Status | Phases 0–3 done. Phase 4 partly done (team login is merged; LAN access is on a branch that is not merged). |
| Works well | Text to image, image to image, image to video, gallery, live progress, team login |
| Stack | Vite + React 19 + TypeScript (frontend) · Node + Express 5 + TypeScript (backend) · MongoDB Atlas · ComfyUI 0.39 |
| Hardware | "System 1": Windows 11, RTX 5090 (32 GB VRAM), 128 GB RAM |
| Size | About 1,300 lines of backend TS, about 1,500 lines of frontend TSX, about 1,300 lines of CSS |
| Dependencies | Very few: backend express, mongodb, dotenv; frontend react, react-dom |
| Tests / lint | None yet. tsc --noEmit passes for both backend and frontend. |
Browser (React, port 5180)
│ only talks to /api
▼
Backend (Express, 127.0.0.1:4000) ──► ComfyUI (127.0.0.1:8188) HTTP + one WebSocket for progress
│
├──► MongoDB Atlas jobs + users (optional: generating still works when it is offline)
├──► data/outputs/ finished .png / .mp4 files
└──► data/inputs/ kept copies of every start image (FRG-25)
Main design ideas
-
Presets are folders.
presets/<id>/workflow.json(a ComfyUI API export) andpreset.json(which user inputs map to which node and field). A new preset needs no code change. OptionalpromptGuideinpreset.json: how to write a good prompt for that model. The chat gives it to Claude. OptionalsendToLabelinpreset.json: the name in the Send to menu (for exampleAnimate). Without it, the preset name.Optional fields (preset schema v2, FRG-32). Old presets without them still load. A bad value writes one warning in the backend log, and the preset loads with the default.
Field Example What it does model"Z-Image-Turbo"The name of the model the preset uses. Chat tells Claude the model. licence"Apache 2.0"Shown as a tooltip on the badge. clientWork"yes","no","conditions"Badge on the preset picker, board cards and gallery detail: Client work OK, Research only, Check licence. Running a Research only preset asks once. licenceNote"Credit the author."One short line. Shown next to Check licence, and in the Research only question. experimentaltrueA small Test label. stages2Sampler passes (1 to 10, default 1). With 2 or more, progress shows Pass 1 of 2 and one bar for the whole job that never goes back. A pass starts when ComfyUI reports progress from another node. estimatedSeconds270A typical run time (1 to 7200). Shown as "Usually takes about 4 min 30 s". The time estimate (FRG-76) uses it when the preset has no finished jobs yet. audioFit{ "input": "audio", "seconds": 4.875 }Makes the voice fit the video (FRG-85). secondsis the real video length of the workflow. A longer voice is cut and the start of the job answers with a warning; a shorter one gets silence at the end. The job savesaudioOriginal,audioUsedandaudioCut.On inputs:
min,maxandstep(kindnumberonly), andoptions(kindnumberortext), for example"options": [{ "value": 768, "label": "Small" }, { "value": 1024, "label": "Medium" }]. The backend refuses other values with a clear message, and chat gets the same limits in its tool. Up to 4 options show as buttons, more as a dropdown. A number input withoutmin/maxkeeps the old rule: 1 to 4096. -
The browser never talks to ComfyUI. ComfyUI has no login and its custom nodes can run code on the PC.
-
No second queue. ComfyUI's own queue is used. The backend only tracks status.
-
MongoDB can fail without breaking generation. You see a warning, not a crash.
-
Login is stateless. A signed HMAC cookie means logged-in users keep working when Atlas is down.
| File | What it does |
|---|---|
server.ts |
Routes, login check, static file serving for outputs, JSON error handling |
jobs.ts |
Job routes. POST /api/jobs: check inputs → save as queued → queue in ComfyUI → answer at once. A 1 s checker follows unfinished jobs (running → copy output → done / failed), saving each change. At start it picks up jobs left queued / running. |
drafts.ts |
Board routes (FRG-22). A card is a job with status draft, a title and a column (idea / ready). Run turns the same record into a queued job. A card whose job fails or is cancelled goes back to Ready with the reason. Anyone can run a card; only who made it can edit, move or delete it (FRG-28). Needs MongoDB. |
chat.ts |
Chat with the Claude API (FRG-23). Conversations in MongoDB (conversations, only the owner can read them), replies streamed with SSE, newest 40 messages sent, 8,000 characters per message, token counts per reply (also in chat_usage), optional daily token limit. The API key never leaves the backend. Each conversation belongs to a project; older ones get "Unsorted" at start, and cards Claude makes go into the conversation's project (FRG-42). Needs MongoDB. |
projects.ts |
Project routes (FRG-39). A project is a named container for jobs, board cards and chats. Names are 1 to 60 characters and unique, not counting capital letters (unique index in projects). Everyone can see and use every project. No delete, only archive. The backend makes the "Unsorted" project at start if it is missing; it cannot be renamed or archived. Needs MongoDB. |
chatPrompt.ts |
The chat system prompt. Edit it to change how Claude answers. |
chatTools.ts |
Chat tools (FRG-24): one create_card_<preset id> tool per preset folder, built for every message from preset.json (no node ids). A tool call is checked like /api/drafts and makes a draft in Ready with the conversation id. It never starts a job. Check errors go back to Claude. At most 8 tool calls per message. Between tool rounds the daily token limit is checked again; when it is passed the reply stops with a note (FRG-28). A start image must be the id of a picture attached in that chat (FRG-25). |
chatImages.ts |
Pictures in chat (FRG-25). Checks each picture against the Claude API limits (PNG/JPEG/WebP, at most 10 MB base64, at most 8000 x 8000 px, at most 20 per message) before any API call. Sends at most the newest 20 pictures and 24 MB of pictures per request; older ones are named in the text. |
inputs.ts |
Start images (FRG-25) and voice files (FRG-84). Keeps every upload in data/inputs, reads the file type (and image size) from the first bytes, and before a job is queued uploads a file again if ComfyUI's input folder lost it. |
lastFrame.ts |
Last frame of a video job (FRG-53). Runs ffmpeg (FFMPEG_PATH in backend/.env) with an argument list, never a shell string, on a video in data/outputs, and saves the last frame as a PNG start image in data/inputs. |
chain.ts |
"Make a longer video" (FRG-56). POST /api/chains saves one chain job (state in job.chain: total length, clip count, one prompt per clip, clip job ids), then makes the clips one after the other as normal jobs, each from the last frame of the one before (parentJobId, chainRootId), and joins them into the chain job (FRG-55). Progress state chain ("Clip 2 of 3" plus the clip's own progress). Cancel stops the chain and the clip being made; a new clip only starts inside the job lock. After a restart, recoverJobs hands the chain back here and it carries on. Lengths come from totalLengths in preset.json. |
join.ts |
Chains of extended clips and "Join all" (FRG-55). The chain view lists the clips from the first clip through this one and onwards (newest extension of each), with lengths read by ffmpeg. A join is a normal job (saved first, followed with GET /api/jobs/:id, progress state joining) that runs ffmpeg in the backend, not in ComfyUI: it drops the first frame of every extension clip, joins, and re-encodes to H.264 / yuv420p MP4. One join at a time; at most 10 clips and 60 seconds. Saves sourceJobIds; goes into the clips' project. |
sendTo.ts |
"Send to" (FRG-26). Copies a finished picture into data/inputs and ComfyUI's input folder, then fills the Generate form or makes a Ready card for any preset with an image input. Saves sourceJobId. Also lists the jobs made from a job ("Used in"). |
presets.ts |
List and load presets, check the optional v2 fields (FRG-32), check user inputs (text / seed / number / image / audio, with min / max / step / options), put values into the workflow. Known preset types: image, video, audio (Music), 3d, edit, upscale, avatar; another type loads with a warning and shows under Other (FRG-30) |
comfyui.ts |
ComfyUI client: queue, history polling, upload, view, error parsing. Reads every list in the output node (images, audio, gifs, 3d, ...): any item with filename, subfolder and type is a file (FRG-29) |
outputKinds.ts |
Accepted output file endings, each with a kind (image, video, audio, 3d) and a content type. Picks the one file a job keeps (an audio or 3d preset prefers a non-picture file). An unknown ending fails the job with a message that names it (FRG-29) |
progress.ts |
One WebSocket to ComfyUI (fixed client id, so it also works after a restart). Step x of y per ComfyUI prompt id |
upload.ts |
Raw image upload. Checks the file's first bytes (magic bytes). 20 MB limit. Our own file name. Keeps a copy in data/inputs. Also "Use in chat": copies a finished picture into data/inputs. Also raw voice upload (FRG-84): WAV, MP3 or FLAC by first bytes, 20 MB and 30 seconds, length read with ffmpeg. |
auth.ts |
scrypt password hashes, HMAC session cookie (7 days), requireLogin middleware |
db.ts |
Mongo connection with a 3 s timeout, jobs and users collections, hides the connection string in logs |
setUser.ts |
CLI npm run user -- <name> to create an account or reset a password |
| Method | Route | Login? | Purpose |
|---|---|---|---|
| POST | /api/login |
no | Log in, sets the session cookie |
| POST | /api/logout |
no | Clears the cookie |
| GET | /api/me |
yes | Who is logged in |
| GET | /api/health |
yes | Backend + ComfyUI reachable? |
| GET | /api/presets |
yes | Preset list (node ids stay hidden from the page) |
| POST | /api/upload |
yes | Upload a start image → ComfyUI input folder, and a copy in data/inputs. Answers { name, width, height } |
| POST | /api/upload-audio |
yes | Upload a voice file → ComfyUI input folder, and a copy in data/inputs. Answers { name, seconds } |
| POST | /api/inputs/from-output |
yes | "Use in chat": { imageUrl: "/api/outputs/<name>" } → copied into data/inputs as a new picture. Answers { name, width, height }. Pictures only |
| GET | /api/inputs/* |
yes | Kept start images and chat pictures (same safe static serving as outputs) |
| POST | /api/jobs |
yes | Start a job from a preset. Answers at once with { jobId } (+ warning if MongoDB is offline). Optional projectId (default "Unsorted"; unknown or archived answers 400). With sourceJobId the job goes into the source job's project (FRG-40) |
| POST | /api/send-to |
yes | { jobId, presetId, target: "generate" | "board" }. generate answers { presetId, inputs, sourceJobId } for the form; board makes a Ready card (prompt still to write) and answers { card }. Pictures only |
| GET | /api/jobs/:id/used-in |
yes | Jobs and cards made from this job with "Send to" |
| PATCH | /api/jobs/:id/project |
yes | Move one job or card to another project: { projectId }. Only projectId changes; files stay where they are. Unknown or archived project answers 400 (FRG-43) |
| POST | /api/jobs/:id/cancel |
yes | Cancel a waiting or running job (removed from ComfyUI's queue, or stopped). Only who started it. A finished job answers 409 |
| POST | /api/jobs/:id/last-frame |
yes | Save the last frame of a finished video job as a new start image in data/inputs. Answers { name, width, height }, like /api/upload. Image jobs, unfinished or failed jobs and bad ids answer 400; an unknown job 404; no FFMPEG_PATH 503 (FRG-53) |
| POST | /api/chains |
yes | "Make a longer video": { presetId, inputs, totalSeconds, prompts?, projectId? }. Makes all clips and joins them. Answers at once with { jobId } of the chain job, which becomes the joined video (FRG-56) |
| GET | /api/jobs/:id/chain |
yes | The chain of clips through this clip, in order: { clips: [{ id, presetId, imageUrl, prompt, seconds }], totalSeconds } (FRG-55) |
| POST | /api/jobs/:id/join |
yes | Join the chain from its first clip up to this clip into one video. Answers at once with { jobId }. More than 10 clips or 60 seconds, a clip with no chain, or another join running answer with a plain-English error (FRG-55) |
| GET | /api/jobs/:id |
yes | One job: status (queued / running / done / failed), progress (waiting: jobs ahead / running: step x of y / finishing), imageUrl (the finished file, of any kind), outputKind (image / video / audio / 3d; worked out from the file ending for older jobs), outputFile, error |
| GET | /api/board |
yes | Board cards: all drafts, all waiting / running jobs (with progress), and the newest 30 finished jobs that started as a card. ?projectId=<id>: only that project's cards (FRG-42) |
| POST | /api/drafts |
yes | Add a card: presetId, inputs, title, column (default idea), projectId (default "Unsorted"). Inputs and project are checked like /api/jobs. Cards from chat go to "Unsorted"; Send to cards go into the source job's project (FRG-40) |
| PATCH | /api/drafts/:id |
yes | Change title, inputs or column. Only while it is a draft, and only by who made it (others get 403) |
| DELETE | /api/drafts/:id |
yes | Delete a card. Only while it is a draft, and only by who made it (others get 403) |
| POST | /api/drafts/:id/run |
yes | Check the inputs again, then queue the card like /api/jobs. Answers { jobId } |
| GET | /api/projects |
yes | Projects that are not archived, sorted by name. "Unsorted" is always there. ?counts=1 adds jobCount (jobs, not counting cards that have not run) (FRG-43). ?summary=1 adds the tile details: jobCount, chatCount (all users), cardCount (cards not run yet), lastActivityAt (newest job, card or chat, else the project's createdAt) and cover (address of the newest finished picture, or null) (FRG-49) |
| POST | /api/projects |
yes | New project: { name }. Empty, longer than 60 characters, or already used (any capital letters, also archived ones) answers 400 |
| PATCH | /api/projects/:id |
yes | Rename or archive: { name?, archived? }. "Unsorted" answers 400 |
| GET | /api/chat/status |
yes | { ready: true }, or { ready: false, error: "Chat is not set up..." } |
| GET | /api/chat/conversations |
yes | My conversations (id, title, projectId, dates), newest first. ?projectId=<id>: only that project's (FRG-42) |
| POST | /api/chat/conversations |
yes | New conversation. Optional title (default "New chat"; the first message becomes the title) and projectId (default "Unsorted"; unknown or archived answers 400) |
| GET | /api/chat/conversations/:id |
yes | One of my conversations with its messages. Someone else's answers 404 |
| PATCH | /api/chat/conversations/:id |
yes | Rename: { title }, and/or move to another project: { projectId } (unknown or archived answers 400, FRG-43) |
| DELETE | /api/chat/conversations/:id |
yes | Delete one of my conversations |
| POST | /api/chat/:conversationId/messages |
yes | { text, images } (max 8,000 characters; images are picture ids from /api/upload, checked against the Claude limits). Streams the reply as SSE: text events, card events (a board card Claude made), then done (the saved reply with token counts and cards) or error |
| GET | /api/jobs |
yes | Job history, newest first. ?limit= (default 50, max 200), ?before=<createdAt> (next page), ?status=done or ?status=active (queued + running), ?mine=1 (only my jobs), ?projectId=<id> (only jobs in this project) |
| GET | /api/outputs/* |
yes | Finished files, each with the content type of its ending (for example .glb → model/gltf-binary) |
App.tsx: login gate, projects state (FRG-41: projects from/api/projects, choice kept inlocalStorageasteczoforge.project, "Unsorted" if none or gone; Generate sends itsprojectId, Gallery shows only its jobs), health check, page switch by URL hash (FRG-46: Chat is the home page at#/,#/chat,#/chat/<conversation id>;#/generate,#/gallery,#/gallery/<job id>,#/board,#/projects).AppShell.tsx(FRG-46, Tailwind + lucide-react): left sidebar with logo, New chat, page links, project picker ("+ New project" at the end), the Recent chats, and the user menu with Log out; Backend/ComfyUI status box at the top right.RecentChats.tsx(FRG-47): the chats of the chosen project in the sidebar, newest first. New chat makes a chat in the project and opens it; on hover Rename, Move to project, Delete (asks first). The list loads again after these, after a project change, and when the Chat page says a chat was made or got its title.PresetBadges.tsx: the licence badge and Test label of a preset, the Research only question, and the estimated time text (FRG-32).Generate.tsx: three columns. Sidebar (categories; Avatar is one of them), Create panel (image upload, prompt, preset picker, Advanced Settings (closed until clicked), Generate), Result panel (large view, download, full screen). It starts empty; older results are in the Gallery.Gallery.tsx: grid of finished jobs → detail view (prompt, maker, preset, seed, size, time) with Use again and Download. Jobs with a start image show Start image → Result. Pictures have Use in chat (attach to the last open chat) and Send to. Shows Made from (link to the start picture's job) and Used in (links).SendToMenu.tsx: the Send to menu (FRG-26) on picture results: result panel, gallery detail, Done board cards and chat cards. One item per preset with an image input (sendToLabelor the name), each with Open in Generate and Add to board.ProjectDialogs.tsx:changeProject(rename or archive, used by the Projects page) and Move to project (used by the gallery job view and the sidebar chat list) (FRG-43).ProjectsPage.tsx(FRG-50, Tailwind + lucide-react):#/projects. Every active project as a tile from/api/projects?summary=1: cover picture (or its first letter), name, files, chats and cards, last activity ("2 days ago"); the current project is highlighted. A click makes it current and opens Chat. The tile menu (…) renames or archives; "Unsorted" has no menu. Create project opens a small dialog and shows the backend's error. MongoDB offline: the backend's message instead of tiles. (Replaces the "Manage projects" dialog.)Board.tsx: four columns (Idea, Ready, Generating, Done). Add / Edit / Delete / Generate on each card, drag and drop between Idea and Ready, Generate all Ready, progress on generating cards, Open and Copy as new card on done cards. Shows only the chosen project's cards; new cards go into it (FRG-42).Chat.tsx: the messages in a centred column and a floating input box at the bottom (FRG-48: Tailwind + lucide-react; the box grows up to 250 px, Enter sends, Shift+Enter makes a new line; empty space under the last message keeps it above the box). The list of chats is in the sidebar (FRG-47). The reply shows while it streams; Stop ends it and keeps the text so far. Simple markdown (paragraphs, lists, bold, code) without a package. Pictures: Attach button, paste, or drag and drop; shown as thumbnails. New chats go into the chosen project. A link to a chat or job of another project switches the picker to that project (FRG-42).ChatCard.tsx: a card Claude made, inside the chat: title, preset, prompt, Generate (runs the draft, like the board), Edit, Open board. Shows progress while it runs and the image or video when done.Media.tsx:outputKind(job)(one helper for every page) and<Media>:<img>for pictures,<video>for videos, for audio the browser player (play, seek, volume, no autoplay) with file name, length and Download, or in grids a tile with a music icon and a small play button (FRG-30), and for 3d a simple tile with an icon, the file name and Download (viewer comes later). Only one sound plays at a time. Send to and Use in chat are only for pictures.Logo.tsx,Login.tsx,styles.css: the older pages' look. The colour variables instyles.cssare the slate and teal set of the sidebar and chat (FRG-51). Icons come fromlucide-react(FRG-51 replaced the hand-drawnIcons.tsx).tailwind.cssloads Tailwind and putsstyles.cssin a layer below Tailwind's utilities.
| Preset | Model | Inputs | Speed (measured) |
|---|---|---|---|
Text to Image (text-to-image-basic) |
Z-Image-Turbo bf16, 8 steps | prompt, seed, width, height (default 1024², 512 to 2048, multiples of 16) | about 2.7 s warm, 7.9 s cold |
Image to Image (image-to-image-basic) |
Z-Image-Turbo, denoise 0.8, about 1 MP | start image, prompt, seed | a few seconds |
Image to Video (image-to-video-basic) |
Wan 2.2 14B I2V fp8 (high + low noise), 20 steps | start image, motion prompt, seed | about 266 s for 5 s of video, 640², 16 fps. Peak VRAM about 29–31 GB |
- ✅ Preset picker built from
preset.json, grouped by category (Image / Image to Video, …) - ✅ Image upload with a preview, file type and size checks in the browser and in the backend
- ✅ Live progress: queue position ("2 jobs ahead"), step progress bar, "Saving the result…"
- ✅ Random seed by default, or a fixed seed you type, with a shuffle button
- ✅ Result viewer with download and full screen. Videos autoplay, loop and are muted.
- ✅ Gallery with video tags, a detail view and "made by"
- ✅ Use again: fills the Generate form from a gallery job
- ✅ Every job is saved to MongoDB (also failed ones, with the error), including the seed that was used
- ✅ Team login (usernames + scrypt hashes in Mongo, accounts made by CLI, no sign-up page)
- ✅ Clear, plain-English error messages, for example "ComfyUI is not running", a missing model from
node_errors, or a file that is too big - ✅ One-click start:
start-teczoforge.batstarts ComfyUI, waits for it, then starts the backend and frontend and opens the browser
Severity: 🔴 fix soon · 🟠 important before team or remote use · 🟡 improvement · 🟢 good as is
| # | Sev | Finding | Suggestion |
|---|---|---|---|
| S1 | 🔴 | backend/.env has no SESSION_SECRET, so a random key is made at every start and everyone is logged out on each backend restart (and tsx watch restarts on every save). |
Add a 64-hex-character secret to backend/.env. |
| S2 | 🟠 | No limit on login attempts. Passwords can be guessed again and again (scrypt slows it down but does not stop it). It matters once the app is on the LAN or Tailscale. | A simple in-memory counter per IP and per username, for example 5 tries then a 1 minute wait. |
| S3 | 🟠 | Sessions cannot be revoked. Log out only deletes the cookie in the browser. A copied cookie works for 7 days, also after a password reset. | Store a sessionVersion on the user and put it in the cookie. A password reset increases it. Or keep a small session list in Mongo. |
| S4 | 🟠 | The cookie has no Secure flag. That is fine on http://localhost, but needed if Tailscale HTTPS (tailscale serve) is used. |
Set secure when the request comes over HTTPS. |
| S5 | 🟡 | No user roles. Every user can see every job, and there is no admin view (users, disk, queue). | Add a role: "admin" field later for delete and user management. |
| S6 | 🟡 | GET /api/progress/:jobId does not check the format of the id (harmless: it is only a map lookup). |
Use the same UUID check as /generate. |
| S7 | 🟢 | Good: ComfyUI is on localhost only. Preset ids are checked against a pattern (no path tricks). Unknown input keys are refused. Image names are checked. Uploads are checked by magic bytes and renamed. Secrets are hidden in logs. The same error is shown for a wrong user and a wrong password. HttpOnly + SameSite=Lax cookie. Timing-safe compares. | — |
| # | Sev | Finding | Suggestion |
|---|---|---|---|
| R1 | 🔴 | /api/generate keeps one HTTP request open for the whole job (videos take about 4.5 min, the limit is 20 min). If the user refreshes, closes the tab, or the network drops, the page loses the job. The backend still finishes and saves it, but the user only sees it later in the gallery. |
Make it async: POST /api/jobs answers at once with a job id. The page then polls /api/jobs/:id (or uses SSE). This also makes a "My running jobs" view possible. |
| R2 | 🔴 | Jobs are written to Mongo only at the end. A backend restart during a job means: ComfyUI still makes the file, but it is never copied to data/outputs and never recorded. |
Insert a queued record first, then update it to running / done / failed. When the backend starts, check ComfyUI /history for jobs that were still open. |
| R3 | 🟠 | There is no Cancel. POST /interrupt and removing from the queue are not used. A wrong 5-minute video blocks the GPU for everyone. |
A Cancel button: /interrupt if it is running, POST /queue {delete:[id]} if it is waiting. |
| R4 | 🟠 | Start images are not kept by the app. They live only in ComfyUI's input/ folder. "Use again" for image-to-image or image-to-video breaks if that folder is cleaned, and the gallery detail does not show the start image. |
Copy uploads to data/inputs/, save the path on the job, and show "Start image → Result" in the detail view. |
| R5 | 🟠 | GET /api/jobs returns every job with no paging. The Generate page loads the whole history just to show 10 thumbnails. It gets slow as the history grows. |
Add ?limit=&before= (cursor on createdAt), add a Mongo index on createdAt, and add infinite scroll in the gallery. |
| R6 | 🟡 | Video progress looks odd: Wan runs two samplers (high noise, then low noise), so the bar goes 0→100% two times. Then the VAE decode and video encode show nothing. | Use preset metadata, for example "stages": 2, or show "Pass 1 of 2". Also show elapsed time and an estimated time left from past durationMs. |
| R7 | 🟡 | Only the first output file is kept (const [file] = …). Batch outputs are dropped. |
Keep all files (outputFiles: []) when batch size is more than 1. |
| R8 | 🟡 | Output file names are presetId-Date.now(). Two jobs that finish in the same millisecond would overwrite each other (rare). |
Add a short random suffix, like uploads already do. |
| R9 | 🟡 | Number inputs only allow 1–4096 for all presets. 4096² on Z-Image may run out of VRAM. Sizes that are not multiples of 16 are allowed. | Let preset.json set min, max, step and options (for example aspect ratio buttons). |
| R10 | 🟡 | The health status is checked only once, right after login. If ComfyUI stops later, the header still says Online. | Check again every 15–30 s, and also after a failed generate. |
| R11 | 🟡 | getUsersCollection() runs createIndex on every login. |
Run it once when the backend starts. |
| R12 | 🟡 | Leftover test file data/outputs/zz-temporary-test-*.png. 36 MB of outputs after a day of tests. There is no disk use view and no way to delete. |
Delete from the gallery (soft delete), and show disk use. |
| # | Sev | Finding | Suggestion |
|---|---|---|---|
| D1 | 🟠 | FRG-15 (LAN access) is pushed but not merged. feat/start-teczoforge-script is also not merged. main stops at FRG-18. |
Review and merge, or close, the open branches. |
| D2 | 🟠 | The app runs with dev servers only (tsx watch + Vite). There is no production build or npm start. (FRG-15 adds a frontend build.) |
A build + start script: Express serves frontend/dist. That means one port and one process for the team. |
| D3 | 🟡 | No tests. | Start with unit tests for buildWorkflow / checkValue and auth (node:test, no new package needed). Add a smoke test that runs all presets against a live ComfyUI. |
| D4 | 🟡 | No lint or format setup. | Prettier only, or tsc --strict as the gate. Keep it small. |
| D5 | 🟡 | start-teczoforge.bat waits forever if ComfyUI never starts, and opens the browser after a fixed 5 s. |
Add a time limit with a clear message, and wait for /api/health before opening the browser. |
| D6 | 🟡 | Logs are only in the console windows. | Add a simple log file or a /api/admin/logs tail for fixing problems from another PC. |
| # | Sev | Finding |
|---|---|---|
| U1 | 🟠 | A lot of the UI is "Soon" placeholders (see section 5). It looks polished, but you cannot tell what is real. Some placeholders should become real soon. |
| U2 | 🟡 | The gallery has no filter, search or sort (by preset, user, image or video, date) and no "only mine". |
| U3 | 🟡 | No negative prompt, no aspect-ratio presets, no batch count (make 4 at once). |
| U4 | 🟡 | You cannot send a result into another preset (for example "Animate this image" → Image to Video, "Upscale this"). This is the core loop of tools like Higgsfield. |
| U5 | 🟡 | There is no team queue view: who is running what, and how long until my job starts. |
| U6 | 🟡 | Images in the result viewer cannot be compared side by side, and there is no keyboard navigation in the gallery. |
| U7 | 🟢 | Good: clear status dots, plain-English errors, the prompt character count matches the backend, "Use again" handles deleted presets. |
| Where | Item | What it needs |
|---|---|---|
| Sidebar → Create | Video (text to video) | A preset with "type": "video" and no image input. Could be Wan 2.2 T2V. |
| Sidebar → Create | Music | A preset with "type": "audio" (FRG-31). Audio results already play everywhere (FRG-30). |
| Sidebar → Create | 3D | A preset with "type": "3d" (for example Hunyuan3D / TRELLIS) and a 3D viewer for .glb files |
| Sidebar → Create | Edit / Inpaint | A preset with "type": "edit" and a mask input (new input kind mask + a brush canvas) |
| Sidebar → Create | Upscale | A preset with "type": "upscale" (for example an ESRGAN model or a tiled diffusion upscale) |
| Create panel | Presets button | A browser for saved prompt and style presets |
| Advanced Settings | Steps selector | A new steps input in preset.json (or Fast / Quality modes) |
The sidebar categories turn on by themselves when a matching preset folder exists. No code change is needed.
- FRG-15: use from other PCs on the same network. (Built on a branch, not merged.)
- FRG-17: remote access with Tailscale, behind the login.
- Later (from
CLAUDE.md): text to video, image editing, upscaling, camera motion presets, consistent characters (LoRA), projects and folders.
- The final video model. Wan 2.2 14B is in use now.
CLAUDE.mdstill lists this as open.
Sorted from quick wins to bigger bets.
Quick wins (mostly no code, or very little)
- "Image to Video (Fast)" preset. The Wan 4-step LightX2V LoRAs are already installed. In the test, turning the switch on gave 32.8 s instead of 266 s (about 8× faster). Export a second workflow with the switch on. It needs no code, only a new folder.
- Add
SESSION_SECRET(S1). - Text to image presets for portrait and landscape (for example 832×1216 and 1216×832).
- Fill Quick Prompts with real starter prompts (a small JSON file).
Medium
5. Async jobs + Cancel + jobs saved first (R1, R2, R3), the biggest gain in reliability.
6. "Send to…" actions on every result: Animate, Upscale, Edit, Use as start image.
7. Gallery filters, paging, "only mine", favorites, soft delete.
8. Keep start images and show them in the detail view (R4).
9. Preset schema v2: min/max/step/options, negative_prompt, batch, aspect choices, estimatedSeconds.
10. Queue view: who is running what, and the estimated wait.
Bigger bets 11. Camera motion presets (dolly in, orbit, crane), the Higgsfield signature feature, using Wan + motion LoRAs. 12. Consistent characters: train or load a LoRA per character, with a "Characters" library to choose from. 13. Projects / folders to group outputs per client or campaign. 14. Prompt helper: a small local LLM that rewrites short prompts into detailed ones. 15. Text to video, 3D, inpaint and upscale presets to light up the "Soon" items. 16. GPU scheduling: show VRAM use, warn before a video job while a big image batch is running, and maybe let urgent jobs go first.
:: Everything at once (System 1)
start-teczoforge.bat
:: Or one at a time
cd backend && npm run dev
cd frontend && npm run dev
:: Add a team account or reset a password
cd backend && npm run user -- <username>
:: Make a SESSION_SECRET for backend/.env
cd backend && npm run secretSet up backend/.env from backend/.env.example (PORT, COMFYUI_URL, MONGODB_URI, SESSION_SECRET,
and for the chat page ANTHROPIC_API_KEY, CLAUDE_MODEL and the optional CHAT_DAILY_TOKEN_LIMIT;
FFMPEG_PATH is the full path to ffmpeg.exe, used to take the last frame of a video).
The chat text goes to the Claude API on the internet. Generation stays on this PC.
For ComfyUI install details, models and measured speeds, see docs/comfyui-setup.md.
ComfyUI and models (not in git). Put the ComfyUI portable build in comfyui\ComfyUI_windows_portable\
and all model files in models\ (same subfolder names ComfyUI uses, plus llm\ for the translate model).
MODELS_DIR in backend/.env can point to another folder, for example on an external drive; empty means models\.
The translate paths (TRANSLATE_MODEL_PATH, TRANSLATE_MMPROJ_PATH) are relative to that folder.
start-teczoforge.bat checks the folder, tells ComfyUI where the models are, and starts ComfyUI minimised.