IMG_TEM_Model

TEM image generation, served over HTTP so projects ask for a picture instead of loading a model themselves. TEM (Transformer Energy Model, paper, code) is a class-conditional ImageNet model: you name one of 1000 classes ("volcano", "golden retriever") and get a 256x256 PNG. There are no text prompts. The service has the same admin page, tokens, rate limits, metrics, settings and update checks as DSC_Laya_Model, CNV_MiniCPM52B_Model, CNV_BitNet2B_Model, TTS_Piper_Model and TTS_Paradee_Model.

The model runs in this process on an NVIDIA GPU (about 4 GB of GPU memory, 6.8 s per image on a laptop RTX 4050). The weights are a 2.4 GB file published by the authors on Google Drive; the admin Download installs it and checks its SHA-256. The TEM code and weights have no stated license: use this for your own projects and ask the authors before anything else. This repository does not contain tem.py; tools/fetch_tem.py downloads it at a pinned commit.

API

Call	Purpose
POST /v1/generate	{"class_id" or "label", "steps", "step_size", "seed"} returns a PNG (image/png, 256x256). Headers: X-Class-Id, X-Class-Label, X-Seed, X-Steps, X-Duration-Ms, X-Cache (hit when the kept image was sent).
GET /v1/classes?q=	{"classes": [{"id", "name"}]}; q keeps the names containing that text.
GET /health	{"status": "ok", "model", "ready", "num_thread", "device"}; ready is true once the model is loaded. No token needed.

Send exactly one of class_id (0 to 999) or label (an ImageNet class name; exact, else the one class that contains the text; 404 if none, 422 if several, with a hint to use class_id). steps is the number of Langevin steps (default 60 as in the paper, at most the Most Langevin steps setting, default 120; more steps are slower). step_size defaults to 500. seed makes the call repeatable: the same request with the same seed gives the same picture (up to GPU rounding); without a seed one is drawn and sent back in X-Seed. Other fields are refused with 422.

Every call starts from fresh noise. The paper's best FID uses a shared buffer of earlier results; this service does not, so the pictures can differ in quality from the paper's (they looked good in tests, with no FID measured).

Cache. The Answers kept per replica setting (SERVICE_CACHE_SIZE, default 32, 0 = off) keeps the last images made with a seed; the same class, steps, step size and seed is sent again without sampling (X-Cache: hit). Calls without a seed are never kept. An image over 1 MB is not kept. The cache is per replica and is emptied when the weights are reloaded.

A token's projects list on the admin page limits the models it may use; there is one model, tem-xl-256.

Admin page

/admin/ (password SERVICE_ADMIN_PASSWORD): tokens with model scope, rate limit and expiry, usage numbers, CPU threads, the Settings card (update check hours, in-flight cap, answers kept, most steps, device, load at start), and the Updates card. The weights are one pinned file, so Check now only finds a missing file and Download new weights downloads it again (after checking the hash). /metrics (admin password) is Prometheus text per token; tem_request_seconds is a histogram of how long each /v1 call took on that replica.

Run

Weights load only from downloaded files. An empty installation still serves /admin/ and /health with ready: false; inference returns 503. Open /admin/ → Updates → Download before the first inference. Checks run online without downloading; only an admin Download replaces weights and reloads replicas.

docker compose --env-file .env.compose up -d --build        # http://<host>:29240, needs the NVIDIA Container Toolkit

Without Docker: run.bat (uv, listens on 127.0.0.1:29240). Setup fetches tem.py; admin Download installs the 2.4 GB checkpoint and the VAE. Time and try it: python tools/bench.py http://127.0.0.1:29240 --save out. See deployment.md.

Tests

uv sync
uv run pytest

tests/test_real_model.py runs the real model and is skipped without _tem_code/, the checkpoint (SERVICE_MODELS_DIR) and a GPU.

Limits worth knowing

• ImageNet classes only, 256x256 only (the authors published no 512 checkpoint). No text prompts, no editing.
• One generation at a time per replica (batching four images on a 6 GB card was slower per image, so there is none); extra calls wait, and more than the in-flight cap get 429.
• No streaming. The reply is the finished PNG.
• 60 steps took about 7 s on a laptop RTX 4050; a CPU takes minutes.
• The rate limit is per replica.

Try it from the admin page

Open /admin/ and use Try it to call the public API with a service token held only in memory. Runs use normal token scopes, limits and caches; the panel shows HTTP status, latency, responses, cancellation and a curl equivalent with a token placeholder. See the web UI guide.

Generate sends exactly one ImageNet class label or class ID to POST /v1/generate, with steps, step size and an optional seed. TEM does not accept a freeform image prompt. The PNG result is shown inline with a download. Only seeded calls are cached; a token's projects list limits the model name tem-xl-256.

Documentation

• Purpose: input, output and what it is for
• Deployment
• Changelog
• Docker on another machine
• Docker on this PC
• Run without Docker
• Test it from the command line

Admin pages and navigation

Open /admin/ and sign in. Overview opens first; each sidebar destination has its own URL fragment and supports browser Back/Forward and direct links:

• #overview: health, model readiness, replicas, active tokens and request counts.
• #playground: service-specific request forms and streamed or media responses.
• #tokens: create, edit, expire and revoke scoped access tokens.
• #runtime: CPU threads where supported, replica information and live/restart settings.
• #updates: check versions and explicitly download weights.
• #connections: caller addresses and public endpoints.
• #models and #voices: project models and cloned voices, shown only when supported.

Collapse/Expand switches between labeled sidebar and icon rail on desktop and mobile. The non-secret sidebar preference is saved per service in this browser. Navigation preserves form drafts and active Playground requests; sign-out still clears credentials and cancels work. Page names appear once in the main header. No preview data or cross-service selector ships in the real admin UI. Existing authentication, quotas, settings precedence and explicit-download policy remain in effect.