Downloads · 30 days
0
adhikjoshi/qie2511-pico-lora-qie2511_pico_10k_2ksteps_v1
qie2511-pico-lora-qie2511_pico_10k_2ksteps_v1 is a machine learning model from adhikjoshi. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
Research and experiment scaffold for training an open-source image editing model toward Nano Banana-style output quality.
Downloads · 30 days
0
Access
Public
Updated May 9, 2026
Repo size
890 MB
Likes
1
Public
Click a slice to open those files.
.safetensors885 MB · 99%
From the Hugging Face model README
Research and experiment scaffold for training an open-source image editing model toward Nano Banana-style output quality.
Last researched: 2026-05-05.
Use Qwen/Qwen-Image-Edit-2511 as the primary base model.
Why:
FLUX.1-Kontext-dev, which is strong but non-commercial.Fallbacks:
Qwen/Qwen-Image-Edit-2509 if 2511 trainer compatibility blocks progress.stepfun-ai/Step1X-Edit-v1p2 as an Apache-2.0 comparison baseline and possible low-VRAM LoRA fallback.black-forest-labs/FLUX.1-Kontext-dev only as a non-commercial benchmark/reference model unless license clearance is obtained.BAAI/OmniGen2 as a broad any-to-any baseline, not the first training target.Before commercial training or release, resolve licensing around Nano Banana-generated teacher data.
Google Gemini API terms include a restriction against using the services to develop competing models. Public Nano Banana-derived datasets also vary:
nano-banana: MIT, but image-only and not enough for instruction-edit learning.The same issue applies to GPT Image-derived datasets. OpenAI's current terms prohibit using Output to develop models that compete with OpenAI. Treat public GPT Image/GPT-4o generated datasets as research-gated until legal review clears the exact use case.
OpenAI now lists gpt-image-2 as the current state-of-the-art GPT Image model with snapshot gpt-image-2-2026-04-21. I did not find a strong public gpt-image-2 training dataset yet.
UCSC-VLAA/GPT-Image-Edit-1.5M is large, but we are marking it as bad data for this project. Do not use it for LoRA training unless a future inspection finds a clean curated subset. The current plan is to avoid it because low-quality synthetic edit pairs can overtrain the LoRA and degrade Qwen's native instruction following, identity preservation, and local-edit behavior.
Rent 1x H100 80GB for the first serious Qwen LoRA run. 1x A100 80GB is acceptable if it is significantly cheaper.
Why:
Qwen/Qwen-Image-Edit-2511, a 20B-class image edit model.Practical sizing:
| Goal | GPU |
|---|---|
| 5k smoke LoRA | 1x H100 80GB preferred; 1x A100 80GB acceptable |
| 10k ablation | 1x H100 80GB |
| 50k curated run | 1x H100 80GB if 12-30h is acceptable; 4x H100 80GB for roughly 4-10h |
| 100k curated run | 1x H100 80GB if 24-60h is acceptable; 4x H100 80GB for roughly 8-20h |
| Step1X fallback LoRA only | 1x RTX 4090 24GB or better |
Avoid starting Qwen-Image-Edit-2511 LoRA on 24GB unless budget forces it. Low-VRAM/offload training can be 3-8x slower and is more operationally fragile. 1x A6000 48GB is better than a 4090 for memory, but still not as comfortable as an 80GB card. 2x RTX 4090 can work, but sharding/orchestration overhead may cost more time than renting one H100.
Minimum useful host:
eval_v1 before training.Last updated: 2026-05-09T03:29:25Z
This table is updated by the checkpoint benchmark loop. Composite is the local frozen-eval proxy score, not an external arena Elo. Use it to compare this LoRA against the same run's base Qwen 2511 result.
HF repo: adhikjoshi/qie2511-pico-lora-qie2511_pico_10k_2ksteps_v1, adhikjoshi/qie2511-pico-lora-qie2511_pico_smoke_5k_1ksteps_v1
| Run | Epoch | Step | Phase | Composite | Delta vs base | Status | Quality movement | LoRA / HF artifact | Benchmark |
|---|---|---|---|---|---|---|---|---|---|
| qie2511_pico_10k_2ksteps_v1 | 0.0000 | 0 | base | 56.917 | 0.000 | base | frozen base reference | base | qie2511_pico_10k_2ksteps_v1/benchmarks/base/result.json |
| qie2511_pico_10k_2ksteps_v1 | 0.1167 | 500 | checkpoint | 56.841 | -0.076 | regressed | composite -0.076; target SSIM -0.0003; PSNR -0.05dB; target MAE 0.0014 higher; source SSIM +0.0004 (more source-like) | qie2511_pico_10k_2ksteps_v1/checkpoints/qie2511_pico_10k_2ksteps_v1_step_000000500.safetensors | qie2511_pico_10k_2ksteps_v1/benchmarks/step_500/result.json |
| qie2511_pico_10k_2ksteps_v1 | 0.2333 | 1000 | checkpoint | 56.670 | -0.247 | regressed | composite -0.247; target SSIM -0.0000; PSNR -0.18dB; target MAE 0.0063 higher; source SSIM +0.0011 (more source-like) | qie2511_pico_10k_2ksteps_v1/checkpoints/qie2511_pico_10k_2ksteps_v1_step_000001000.safetensors | qie2511_pico_10k_2ksteps_v1/benchmarks/step_1000/result.json |
| qie2511_pico_10k_2ksteps_v1 | 0.3500 | 1500 | checkpoint | 56.893 | -0.024 | regressed | composite -0.024; target SSIM -0.0009; PSNR -0.02dB; target MAE 0.0007 lower; source SSIM +0.0017 (more source-like) | qie2511_pico_10k_2ksteps_v1/checkpoints/qie2511_pico_10k_2ksteps_v1_step_000001500.safetensors | qie2511_pico_10k_2ksteps_v1/benchmarks/step_1500/result.json |
| qie2511_pico_smoke_5k_1ksteps_v1 | 0.0000 | 0 | base | 54.347 | 0.000 | base | frozen base reference | base | qie2511_pico_smoke_5k_1ksteps_v1/benchmarks/base/result.json |
| qie2511_pico_smoke_5k_1ksteps_v1 | 0.0521 | 250 | checkpoint | 54.098 | -0.250 | regressed | composite -0.250; target SSIM -0.0021; PSNR -0.15dB; target MAE 0.0040 higher; source SSIM -0.0039 (less copied) | qie2511_pico_smoke_5k_1ksteps_v1/checkpoints/qie2511_pico_smoke_5k_1ksteps_v1_step_000000250.safetensors | qie2511_pico_smoke_5k_1ksteps_v1/benchmarks/step_250/result.json |
| qie2511_pico_smoke_5k_1ksteps_v1 | 0.1041 | 500 | checkpoint | 54.604 | 0.256 | flat/slight lift | composite +0.256; target SSIM +0.0064; PSNR -0.02dB; target MAE 0.0013 higher; source SSIM +0.0042 (more source-like) | qie2511_pico_smoke_5k_1ksteps_v1/checkpoints/qie2511_pico_smoke_5k_1ksteps_v1_step_000000500.safetensors | qie2511_pico_smoke_5k_1ksteps_v1/benchmarks/step_500/result.json |
| qie2511_pico_smoke_5k_1ksteps_v1 | 0.1562 | 750 | checkpoint | 55.122 | 0.775 | flat/slight lift | composite +0.775; target SSIM +0.0152; PSNR +0.11dB; target MAE 0.0024 lower; source SSIM +0.0102 (more source-like) | qie2511_pico_smoke_5k_1ksteps_v1/checkpoints/qie2511_pico_smoke_5k_1ksteps_v1_step_000000750.safetensors | qie2511_pico_smoke_5k_1ksteps_v1/benchmarks/step_750/result.json |
| qie2511_pico_smoke_5k_1ksteps_v1 | 0.2082 | 1000 | checkpoint | 55.150 | 0.802 | flat/slight lift | composite +0.802; target SSIM +0.0164; PSNR +0.09dB; target MAE 0.0020 lower; source SSIM +0.0103 (more source-like) | qie2511_pico_smoke_5k_1ksteps_v1/checkpoints/qie2511_pico_smoke_5k_1ksteps_v1_step_000001000.safetensors | qie2511_pico_smoke_5k_1ksteps_v1/benchmarks/step_1000/result.json |
Reference targets for Nano Banana, GPT Image, and Qwen arena scores are tracked in configs/benchmark_references.yaml. Those scores are not numerically comparable with the local proxy score above.
The current qie2511_pico_10k_2ksteps_v1 run is a distribution-mismatch diagnostic, not a clean scale result. The intended 10k Pico slice was downshifted to 4,286 accepted train rows after Flickr rate limiting, and the frozen eval split is fully out-of-distribution by edit_type: manifest_distribution_overlap.json reports eval_share_absent_from_train = 1.0 with recommendation rebalance_before_training.
Do not use this run alone to judge Pico-Banana quality. The next serious run should use Open Images local source mapping, stratified train/eval sampling by edit type, and manifest overlap checks before training.
Step 1000 was also tested with Qwen's edit-plus multi-image path using two input images per prompt. The run completed successfully on the Blackwell host, so future checkpoints can include a separate multi-image smoke eval.
artifacts/benchmarks/qie2511_pico_smoke_5k_1ksteps_v1/multi_image_step_1000/result.jsonartifacts/benchmarks/qie2511_pico_smoke_5k_1ksteps_v1/multi_image_step_1000/contact_sheet.jpgruns/qie2511_pico_smoke_5k_1ksteps_v1/staged_checkpoints/qie2511_pico_smoke_5k_1ksteps_v1_step_000001000.safetensorsThis is a visual/API probe only, not part of the single-image composite score. The next-source reference mode proves multi-image input works, but several examples show reference-image leakage because the second image is unrelated to the requested edit. Future multi-image eval should use controlled references such as style, identity, object, or layout exemplars.