Text-to-Image
Diffusers
Safetensors
flow-matching
pyramidal-unet
256x256
ettin-150m
from-scratch
ra
udit
Instructions to use AiArtLab/Ra with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Diffusers
How to use AiArtLab/Ra with Diffusers:
pip install -U diffusers transformers accelerate
import torch from diffusers import DiffusionPipeline # switch to "mps" for apple devices pipe = DiffusionPipeline.from_pretrained("AiArtLab/Ra", dtype=torch.bfloat16, device_map="cuda") prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k" image = pipe(prompt).images[0] - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- Draw Things
- DiffusionBee
Card: compact
Browse files
README.md
CHANGED
|
@@ -1,27 +1,16 @@
|
|
| 1 |
---
|
| 2 |
license: mit
|
| 3 |
pipeline_tag: text-to-image
|
| 4 |
-
tags:
|
| 5 |
-
- text-to-image
|
| 6 |
-
- flow-matching
|
| 7 |
-
- diffusion-transformer
|
| 8 |
-
- pyramidal-unet
|
| 9 |
-
- 256x256
|
| 10 |
-
- ettin-150m
|
| 11 |
-
- from-scratch
|
| 12 |
-
- ra
|
| 13 |
-
- udit
|
| 14 |
datasets:
|
| 15 |
- LucasFang/FLUX-Reason-6M
|
| 16 |
---
|
| 17 |
|
| 18 |
# Ra
|
| 19 |
|
| 20 |
-
**Ra** is a **256×256 text-to-image** model trained from scratch: 177.3M parameters
|
| 21 |
-
|
| 22 |
-
|
| 23 |
-
flow head (23.9M). Named after the sun god because the body is a pyramid of widths: 192 channels
|
| 24 |
-
where the grid is dense, 1536 where only 16 tokens remain.
|
| 25 |
|
| 26 |
| | |
|
| 27 |
|---|---|
|
|
@@ -31,19 +20,16 @@ where the grid is dense, 1536 where only 16 tokens remain.
|
|
| 31 |
| training | FLUX-Reason-6M, one RTX 4080 |
|
| 32 |
| licence | MIT |
|
| 33 |
|
| 34 |
-
**Training is still running
|
| 35 |
-
restarted at 1e-4 on step 192 000, so
|
| 36 |
|
| 37 |
## Examples
|
| 38 |
|
| 39 |
-
Step 212 000, 256 px, 16 body steps, 8 head steps, cfg 2.5, seed 0
|
|
|
|
| 40 |
|
| 41 |
<img src="docs/img/samples.png" width="100%"/>
|
| 42 |
|
| 43 |
-
Reading order: red fox in fresh snow · sailboat at sunset · portrait of an elderly fisherman ·
|
| 44 |
-
sunflowers under a stormy sky · blue morpho on an orchid · white butterfly with orange tips ·
|
| 45 |
-
dark swallowtail on white · monarch on a pink flower.
|
| 46 |
-
|
| 47 |
The butterfly domain it started from, before the FLUX-Reason stream (step 20 003, earlier weights):
|
| 48 |
|
| 49 |
<img src="docs/img/samples_butterflies_20k.png" width="100%"/>
|
|
@@ -56,11 +42,8 @@ python3 src/sample.py --prompt "A macro photograph of a blue morpho butterfly."
|
|
| 56 |
|
| 57 |
## Limitations
|
| 58 |
|
| 59 |
-
- Animal anatomy fails (see the fox), hands fail; detail finer than 8 px is invented by the
|
| 60 |
- Loose caption adherence, no glyphs, 256² only. Research artifact — no safety tuning, no filter.
|
| 61 |
|
| 62 |
-
|
| 63 |
-
|
| 64 |
-
`transformer/` weights (step 212 000) · `vae/` f8c32-ae · `text_encoder/` Ettin-150M · `src/` model,
|
| 65 |
-
trainer, sampler · `docs/ARCH.md` architecture, `docs/STATUS.md` plans, `docs/presentation.pdf` (RU).
|
| 66 |
-
MIT for our code and weights; the bundled Ettin-150M is MIT and f8c32-ae is Apache-2.0.
|
|
|
|
| 1 |
---
|
| 2 |
license: mit
|
| 3 |
pipeline_tag: text-to-image
|
| 4 |
+
tags: [text-to-image, flow-matching, pyramidal-unet, 256x256, ettin-150m, from-scratch, ra, udit]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 5 |
datasets:
|
| 6 |
- LucasFang/FLUX-Reason-6M
|
| 7 |
---
|
| 8 |
|
| 9 |
# Ra
|
| 10 |
|
| 11 |
+
**Ra** is a **256×256 text-to-image** model trained from scratch: 177.3M parameters — a pyramidal
|
| 12 |
+
U-transformer (143.3M) flowing in an 8×/32-channel latent, a frozen Ettin-150M text encoder with a
|
| 13 |
+
learnable layer-slice mixer (10.0M), and the AE decoder retrained as a pixel-space flow head (23.9M).
|
|
|
|
|
|
|
| 14 |
|
| 15 |
| | |
|
| 16 |
|---|---|
|
|
|
|
| 20 |
| training | FLUX-Reason-6M, one RTX 4080 |
|
| 21 |
| licence | MIT |
|
| 22 |
|
| 23 |
+
**Training is still running** — the bundled `transformer/` is a snapshot of step **212 000** (lr was
|
| 24 |
+
restarted at 1e-4 on step 192 000, so it is mid-recovery and will be replaced).
|
| 25 |
|
| 26 |
## Examples
|
| 27 |
|
| 28 |
+
Step 212 000, 256 px, 16 body steps, 8 head steps, cfg 2.5, seed 0 — fox · sailboat · fisherman ·
|
| 29 |
+
sunflowers · blue morpho · white butterfly · swallowtail · monarch:
|
| 30 |
|
| 31 |
<img src="docs/img/samples.png" width="100%"/>
|
| 32 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 33 |
The butterfly domain it started from, before the FLUX-Reason stream (step 20 003, earlier weights):
|
| 34 |
|
| 35 |
<img src="docs/img/samples_butterflies_20k.png" width="100%"/>
|
|
|
|
| 42 |
|
| 43 |
## Limitations
|
| 44 |
|
| 45 |
+
- Animal anatomy fails (see the fox), hands fail; detail finer than 8 px is invented by the decoder.
|
| 46 |
- Loose caption adherence, no glyphs, 256² only. Research artifact — no safety tuning, no filter.
|
| 47 |
|
| 48 |
+
MIT for our code and weights (`src/`, `transformer/`); bundled Ettin-150M is MIT, f8c32-ae
|
| 49 |
+
Apache-2.0. Architecture notes: `docs/ARCH.md`, Russian slides: `docs/presentation.pdf`.
|
|
|
|
|
|
|
|
|