Same v2 content (padding to 1:1 for whole-image mode) — pushed to the
new filename following the aioli-subgraph_<model> convention that also
applies to the two new Flux.2 Klein and Qwen variants.
In whole-image mode (inpaint_mode=False), the source image is now
automatically padded to a 1:1 square before being sent to nano-banana,
then cropped back to its original dimensions after recompose. This
prevents nano-banana's internal recrop (which it applies to non-square
inputs) and preserves every pixel of the original aspect ratio.
New subgraph nodes (11):
- GetImageSize+ (110): reads ORIGINAL input dims before pad
- ComfyMathExpression 910/911: pad_w_total / pad_h_total
(0 in zone mode, ceil(max(w,h)/8)*8 - dim in whole mode)
- ComfyMathExpression 920-923: pad_left / right / top / bottom
- ImagePadForOutpaint (930): pads with feather=0
- SolidMask (940) + MaskComposite (941): pads the user mask in sync
- ImageCrop+ (950): final crop back to orig dims after composite
Rewiring:
- 100/212/219.destination now read from 930 (padded image)
- 600/603 now read from 941 (padded mask)
- recomposed_image output now routes through 950
Behaviour preservation:
- Zone mode (inpaint_mode=True): all pad values = 0, new nodes are
pure no-ops. Strictly identical to v1.
- target_size applies AFTER the pad, so users can still control
the resolution sent to nano-banana (e.g. 1024 → sends 1024×1024).
- Already-square sources produce no padding (pads = 0).
Note: padding uses neutral grey (0.5), which is stripped off at the
final crop — never appears in the output.
Full 44KB minified JSON with all 30 subgraph nodes, 93 subgraph links,
proxyWidgets for force_square/inpaint_mode/use_mask_blur/multiple/target_size,
and the updated MarkdownNote documenting usage, modes, anti-clamp,
troubleshooting (including LoadImageOutput → Load Image swap for local),
and limitations.
Pure subgraph workflow that reproduces BBoxMultipleFix using only nodes
available on ComfyUI Cloud (no custom install required). Pipes to
nano-banana (Gemini Image) for inpainting.
Features:
- force_square, zone/whole inpaint modes, mask blur toggle
- multiple alignment, target_size resize (0/512/768/1024/1536/2048)
- Anti-clamp: crop never exceeds source bounds
- Alignment enforced after clamp (no multiple drift)
- Whole-image mode uses a full white mask for recompose
- LoadImageOutput can be swapped for Load Image on local installs
/ai.oli