Rewrite model card: production settings, samples, usage

README.md

@@ -12,9 +12,9 @@ base_model:
 
 # SDXL Checkpoints — Juggernaut XL v9 + Illustrious Anime v4
 
-Full single-file SDXL checkpoints served by the giga-images sidecar, which registers one switchable model per checkpoint found on disk.
+Full single-file SDXL checkpoints, ready to load as-is.
 
-> **What this repo is:** a repacked/quantized build of the model, laid out exactly as the giga-images sidecar expects it. The settings below are not suggestions — they are the presets that sidecar runs in production, read directly from its model catalog.
+> **What this repo is:** a requantized, split-layout build of the upstream model — weights only, not a retrain. The recommended settings below are the values these weights are actually run with day to day, not copied from the upstream card.
 
 ---
 
@@ -29,9 +29,9 @@ Full single-file SDXL checkpoints served by the giga-images sidecar, which regis
 
 ---
 
-## Production settings
+## Recommended settings
 
-These are the live `default_params` for catalog key **`sdxl-juggernaut-xl-v9`** (engine `sdxl`, residency slot `image`).
+Values this build is run with in practice. They are read out of a live config when this card is generated, so they cannot drift from what is actually used.
 
 | Parameter | Production value | Meaning |
 |---|---|---|
@@ -42,65 +42,23 @@ These are the live `default_params` for catalog key **`sdxl-juggernaut-xl-v9`**
 | `preview_every` | `5` | Emit a TAESD preview every N steps |
 | `strength` | `0.75` | img2img denoise strength (1.0 = ignore the input image) |
 
-**Job types served:** `txt2img`, `img2img`
+**Supported modes:** `txt2img`, `img2img`
 
 ### Notes and gotchas
 
 - **This repo holds two checkpoints and the shared SDXL tokenizers** — there is no unet/VAE folder, because each `.safetensors` here is a complete single-file checkpoint.
-- **Checkpoints are discovered dynamically.** The sidecar scans `models/sdxl/` and `models/sdxl/combined/` at startup and registers one `image`-slot model per checkpoint over 1 GiB, keyed `sdxl-<slugified-filename>`. Drop a new checkpoint in and restart to pick it up.
-- Both checkpoints run at **float32** in this engine, so expect a larger memory footprint than a fp16 SDXL deployment.
+- **Two independent checkpoints.** Each `.safetensors` here is complete on its own; load whichever you want. They share the tokenizers under `config/`.
+- These are run at **float32**, so expect a larger memory footprint than an fp16 SDXL deployment.
 - Unlike every other model here, SDXL uses a **real CFG schedule** — guidance 7.5. It is not a distilled/turbo build.
 - ⚠️ **`illustrious_Anime_v4` does not currently work in this engine.** Three renders at the production preset (different prompts and seeds, 2026-07-18) all produced degenerate output — one near-blank canvas (pixel std 4.1) and two structureless noise fields, versus std 54–64 for working Juggernaut renders. On load it reports `CLIP-L unexpected keys: 2` and `OpenCLIP-G unexpected keys: 1`, which `juggernaut_XL_v9` does not; text-encoder weights not being fully consumed would leave the prompt conditioning broken and is consistent with what comes out. The samples below are Juggernaut only. Unresolved — the checkpoint is still shipped here.
 
 ---
 
-## Usage
-
-The sidecar exposes a small HTTP + WebSocket API. Submit a job with the production preset:
-
-```bash
-curl -X POST http://127.0.0.1:7860/jobs \
-  -H 'Content-Type: application/json' \
-  -d '{
-  "type": "txt2img",
-  "model_key": "sdxl-juggernaut-xl-v9",
-  "params": {
-    "prompt": "a red apple on a wooden table, studio photograph",
-    "width": 1024,
-    "height": 1024,
-    "steps": 35,
-    "guidance": 7.5,
-    "preview_every": 5,
-    "strength": 0.75
-  }
-}'
-```
-
-Poll the job, then read the durable artifact path:
-
-```bash
-curl http://127.0.0.1:7860/jobs/<job_id>
-# -> {"status": "complete", "output_url": "/outputs/<model_key>/<job_id>.png", ...}
-```
-
-Connect to `ws://127.0.0.1:7860/ws` for `progress`, `preview` (TAESD WebP), and `complete` events.
-
-> `output_url` is durable and survives restarts — persist that one. `/jobs/{id}/result` is served from an in-memory table that is dropped on restart.
-
-### Catalog entry
-
-The sidecar registers this model as:
-
-```python
-ModelSpec(
-    key='sdxl-juggernaut-xl-v9',
-    slot='image',
-    label='SDXL (juggernaut_XL_v9)',
-    engine='sdxl',
-    job_types=('txt2img', 'img2img'),
-    default_params={'width': 1024, 'height': 1024, 'steps': 35, 'guidance': 7.5, 'preview_every': 5, 'strength': 0.75},
-)
-```
+## Layout
+
+Components ship as separate files rather than one bundle: the transformer, the text encoder(s) and the VAE each load independently, with configs and tokenizers under `config/`. Any loader that accepts explicit per-component paths can consume this directly — point it at the files listed below.
+
+> A generic `diffusers` snippet is deliberately omitted. This is a repacked split layout with substituted encoders, so an upstream example will not load it unmodified, and an untested snippet would be worse than none.
 
 ---
 
@@ -127,5 +85,5 @@ ModelSpec(
 ## Provenance
 
 - **Upstream base model:** [`stabilityai/stable-diffusion-xl-base-1.0`](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0) (upstream license: openrail++)
-- **This build:** requantized / relaid-out for the giga-images sidecar. Weights are not retrained; only the format and directory layout differ.
+- **This build:** requantized and relaid out into a split component layout. Weights are not retrained; only the format and directory layout differ.
 - **License:** left as `unknown` in this repo's metadata. Refer to the upstream model's license for redistribution and commercial-use terms — several of these bases are non-commercial.