Axon Image 2 Flash
A trained image stack with a quality ladder: a compact model at 4 steps for drafts, the same stack scaled to a 20B backbone compacted to 4 bits when a draft is not good enough.
Weights of the backbones are not redistributed. This repository ships our trained adapters, the serving engine, and the configuration that binds them to public base models.
Tiers
| tier | backbone | steps | guidance | resident | hardware |
|---|---|---|---|---|---|
flash |
FLUX.2-klein 4B | 4 | 1.0 | ~8 GB fp16 | any 16 GB card, free T4 |
hq |
Qwen-Image 20B + 8-step distillation | 8 | 1.0 | ~41 GB bf16 / ~21 GB W8 | 24-48 GB |
pro |
Qwen-Image 20B, W4 | 30 | 4.0 | ~11 GB | 24 GB L4 |
pro is the same 20B denoiser with the residual quantised to 4 bits and the low-rank branch kept in
16 bits (SVDQuant). Published checkpoints: nunchaku-ai/nunchaku-qwen-image, svdq-int4 / svdq-fp4
at rank 32 and 128. Nothing is pruned; the model is compacted, not deleted.
What it does well
- Legible text. Quoted strings can be locked: the engine renders, reads the result back, scores
it and retries; if no candidate reads correctly the literal is set as real type on a caption bar.
Every response reports
verified,compositedorunverified, so a caller always knows which one they got. - Subject consistency. A
subjectname derives the seed deterministically, so repeat calls reuse the same character with no server state; up to 4 reference images on/v1/images/editsis the stronger lever. - UI, motion-graphics and game framing. Dashboards, app screens, landing pages, lower thirds, title cards, HUD and inventory screens, gameplay look per title.
- Speed as a design constraint, not an afterthought. 4 steps at 1.0 guidance for drafts, with the ladder up to 30 steps only when it earns the latency.
Not a frontier photoreal model: small text in dense layouts can still misread, and fine detail is bounded by the backbones. This is a control-and-speed layer, and the tiers are chosen accordingly.
Adapters
| file | backbone | what |
|---|---|---|
adapters/games-v1.safetensors |
klein 4B | gameplay look across 20 titles |
adapters/games2-v1.safetensors |
klein 4B | gameplay look across 15 further titles, more distinct per-title styling |
Trained by us on our own corpora; both load with pipe.load_lora_weights(repo, weight_name="adapters/<file>").
Usage
from openai import OpenAI
c = OpenAI(base_url=f"{AXON_URL}/v1", api_key=AXON_API_KEY)
r = c.images.generate(model="axon-image-2-flash", size="1536x1024", quality="flash", n=2,
prompt='app dashboard titled "NORTH ARC" with a revenue chart',
extra_body={"text_lock": True, "subject": "north arc"})
Run it on a free T4 (Colab)
Axon_Image_2_Flash_Colab.ipynb is a two-cell notebook that loads the 4B backbone (4-bit per
component on a 12.7 GB-RAM T4, fully resident on L4/A100), serves this same HTTP surface on port
8199 and opens a public tunnel. Cell 3 is a smoke test: one text image, one two-reference edit, one
text-locked render. HF_TOKEN is only needed if your account has to accept the base model's terms.
Deploy serverless with Modal: scripts/deploy_modal.sh for the flash tier, --pro for the W4 20B.
scripts/bench_api.py measures a deployment; scripts/stub_api.py serves the API against a stub
engine so the HTTP contract can be tested without a GPU.
| endpoint | purpose |
|---|---|
POST /v1/images/generations |
prompt, size, n, quality, seed, negative_prompt, refine, subject, text_lock, response_format |
POST /v1/images/edits |
multipart, up to 4 reference images |
GET /v1/images/file/{name} |
fetch a url response |
GET /v1/models, GET /healthz |
capability probe |
Licence and provenance
Backbones are Apache-2.0 (Qwen-Image, FLUX.2-klein); distillation adapters by LightX2V; W4 checkpoints by Nunchaku. Our code and adapters here are Apache-2.0. Base weights are not included and resolve from their publishers.