README.md

---
title: Linarix V2 T2I
emoji: 🎨
colorFrom: indigo
colorTo: purple
sdk: gradio
sdk_version: 6.20.0
app_file: app.py
python_version: 3.11
pinned: false
license: other
short_description: 1024px text-to-image with Linarix-v2 (GDN-2) on ZeroGPU
models:
  - Akrao9/Linarix-v2
  - Qwen/Qwen3.5-4B
  - mit-han-lab/dc-ae-f32c32-sana-1.1-diffusers
---

# Linarix-v2 — Text to Image Demo

Interactive demo for [Akrao9/Linarix-v2](https://huggingface.co/Akrao9/Linarix-v2):
a Boomer v8 flow-matching model with a bidirectional GatedDeltaNet-2 semantic encoder,
substitutive joint-attention refresh layers, Block AttnRes, and a full-attention DDT
velocity decoder. The app reads the resolution and inference defaults from the exported
checkpoint instead of duplicating them here.

**Capabilities:** strong on landscapes, scenic/architectural scenes, **rigid man-made
structures** (lighthouses, locomotives, bridges), and composed portraits. Text
conditioning uses **Qwen3.5-4B layers 9/18/27** with per-layer RMS normalization and
feature concatenation; decode uses **DC-AE f32c32**; sampling uses **STORK-4**. Step
count, CFG scale/rescale, guidance interval, RKG substeps and flow shift all come from
the checkpoint's own metadata, so they are deliberately not restated here — the Status
line under each generation reports what actually ran.

## Space setup

1. Create a new **Gradio** Space on Hugging Face.
2. Upload the contents of this `Space/` folder (or point the Space repo here).
3. In **Settings → Hardware**, select **ZeroGPU** (requires HF PRO for hosting). This allocates a shared **NVIDIA RTX Pro 6000 Blackwell** GPU (48 GB via `@spaces.GPU(size="large")`).
4. **Required secret:** add `HF_TOKEN` (Settings → Secrets) with a token that can read gated `Akrao9/Linarix-v2`. Spaces `preload_from_hub` cannot auth gated/private repos, so the model is downloaded at runtime via `snapshot_download(token=...)`.
5. Restart / Factory rebuild the Space after setting the secret (model + VAE + text encoder download can take a few minutes on first boot).

## Local test

```bash
cd Space
pip install -r requirements.txt
python app.py
```

The `@spaces.GPU` decorator is a no-op locally; use a CUDA machine for full inference.

Pinned deps for ZeroGPU: do **not** pin `torch` or `gradio` in `requirements.txt` — the Space builder injects `torch<=2.11.0` and `gradio[oauth,mcp]==<sdk_version from this README>` into the same pip install, so a gradio pin that differs from `sdk_version` makes the build fail with `ResolutionImpossible`. Include **`torchvision`** (unpinned) for the Qwen3.5 `AutoProcessor`. Other pins: diffusers 0.38.0, transformers 5.9.0, and flash-linear-attention at the git
SHA in `requirements.txt` — a released version tag is not enough, since Block AttnRes
needs the same FLA build the checkpoint was exported against.

`app.py` redirects Hugging Face caches to `/tmp/huggingface` (ZeroGPU `~/.cache` is often read-only).

Set **`python_version: 3.11`** in this README (Spaces default is 3.10). After changing it, **Factory rebuild** the Space so the runtime upgrades.

## Startup warnings (usually harmless)

| Message | Cause | Action |
|---|---|---|
| Python 3.10 below recommended 3.11 | Spaces default until `python_version: 3.11` is applied | Factory rebuild after README change |
| `Triton is not supported… roll back to CPU` | FLA probes Triton at import; on some builders/GPUs it falls back to non-Triton CUDA kernels | Safe if generation works; rebuild on 3.11 |
| `torch.compile is not available in Python 3.10` | FLA training helper stub; **inference does not use torch.compile** | Fixed by Python 3.11 rebuild |
| `local_dir_use_symlinks` deprecated | Older `huggingface_hub` call inside a dependency | Cosmetic; ignore or upgrade hub later |

## Notes

- Uses `@spaces.GPU(size="large", duration=150)` — RTX Pro 6000 Blackwell 48 GB tier —
  for STORK-4 generation ([Tan et al., 2025](https://arxiv.org/abs/2505.24210)).
- Example prompts run generation on click (`run_on_click=True`).
- The CFG scale, rescale and guidance interval defaults come from the exported
  checkpoint; the sliders start at those values. For exact object counts, sweep a
  few seeds.
- VAE and Qwen3.5-4B are fetched from their upstream repos on first startup.