About this model
Aurora XL — Photoreal Base Checkpoint
A general-purpose photoreal base checkpoint — natural skin, believable light and clean anatomy, no LoRA required.
Aurora XL is a full fine-tune built to be your everyday photoreal base: convincing skin micro-texture, physically plausible lighting, and dependable hands and anatomy across portraits, interiors and product scenes. It is a complete checkpoint — load it and generate, or use it as the base for your own LoRAs.
This is a complete, standalone checkpoint. Unlike a LoRA, it needs no separate base model — load it directly as your base and generate.
Model details
| Developed by | @aria-forge |
| Model type | Full fine-tuned checkpoint |
| Base architecture | wan 2.2 (full fine-tune) |
| Parameters | ≈ 2.6 B |
| Precision / format | fp16 · safetensors |
| VAE | baked in |
| Weights | aurora_xl_v2.safetensors — 6.5 GB |
| License | Personal ($59.00) · Commercial ($299.00) |
| Category | photoreal |
Uses
Direct use. A default photoreal base for portraits, lifestyle, interiors and product work — and a strong foundation to train or stack style LoRAs on.
Out of scope. Hard stylisation (anime, painterly, flat illustration) — for those, stack a style LoRA or use a stylised checkpoint; Aurora pulls back toward realism.
Recommended settings
| Setting | Value |
|---|---|
| Sampler | DPM++ 2M Karras |
| Steps | 28–32 |
| CFG scale | 4.5–6 |
| Resolution | 1024×1024 (to 1536 on the long edge) |
photograph of a woman by a rain-streaked window, soft overcast light, natural skin texture, 50mm, shallow depth of field, high detail
The baked VAE means no separate VAE file is needed. Keep CFG moderate — above 7 the skin starts to look plasticky; a light hi-res fix pass at 1.5× is the intended finishing step.
- Strengths — natural skin and portrait lighting; reliable hands and anatomy; interiors and product realism.
- Weaknesses — hard stylised or illustrative looks; extreme close-up macro detail.
Samples
| Window portrait | Interior scene |
|---|---|
Motion sample
A short clip rendered with this listing — click to play the HLS stream.
Training details
A full fine-tune over roughly 90k licensed and owned photographs, aesthetic-scored and captioned in-house. No scraped datasets; face-bearing images are model-released. A second refinement pass used a hand-picked 6k high-aesthetic subset.
Procedure. A full fine-tune of wan 2.2 (not an adapter), fp16, multi-stage: a broad concept pass followed by an aesthetic refinement pass on a hand-scored subset. VAE was fine-tuned and baked into the checkpoint.
Bias, risks and limitations
- Very large checkpoint (~6.5 GB) — plan VRAM and storage accordingly.
- Photoreal bias means stylised prompts under-respond without a LoRA.
- Like all base models it can still miss on complex hand interactions at low step counts.
License & provenance
A full fine-tune on licensed and owned, model-released photography. Purchasing the commercial offer ($299.00) grants a commercial redistribution license for outputs; the personal offer ($59.00) is for non-commercial use. The checkpoint may not be resold or re-uploaded, and merges must credit this model.
Citation
@misc{aria-forge_00000017,
title = {Aurora XL — Photoreal Base Checkpoint},
author = {Aria Forge Studio (@aria-forge)},
year = {2026},
note = {TensorForge marketplace checkpoint, LST-00000017-0000-4000-8000-000000000017}
}