NidAll/Krea-2-BBOX-Turbo-INT8-ConvRot

Krea 2 BBOX Turbo — INT8 ConvRot

INT8 ConvRot quantizations of jimmycarter/krea2-turbo-bbox, the Krea 2 finetune for layout control: multi-panel comics, characters that stay themselves across panels, and verbatim text placed in explicit boxes.

These checkpoints are modified versions of the Krea Model (quantized derivatives of a finetune of Krea 2 Raw). They are not official Krea products and are not endorsed by Krea.

Krea 2 is licensed under the Krea 2 Community License Agreement. For more information, visit https://krea.ai/krea-2-licensing. The full agreement is in LICENSE.md in this repo. Note in particular: commercial use under that license is permitted only below $1M annual revenue (Section 2.3); above that, an Enterprise License from Krea is required.

Files

File	Variant	Source size	INT8 size
krea2-bbox-turbo-int8-convrot.safetensors	Turbo fine-tune	26.28 GB	14,135,120,528 bytes
krea2-bbox-turbo-ema_0.9995-int8-convrot.safetensors	Turbo EMA 0.9995	26.28 GB	14,135,120,552 bytes
krea2-bbox-turbo-ema_0.99975-int8-convrot.safetensors	Turbo EMA 0.99975	26.28 GB	14,135,120,568 bytes

The INT8 size target is the official Comfy-Org krea2_turbo_int8_convrot.safetensors size: these conversions follow that artifact's exact recipe (same quantized layer set, same per-layer ConvRot group sizes), so the sizes align. Measured sizes from the conversion run:

SHA256 of every file: see checksums-out.txt. Which variant to pick: the plain fine-tune follows boxes most precisely. The EMA shadows are smoothed averages of the training weights; they can look slightly cleaner but follow boxes a little less exactly (0.99975 more so than 0.9995).

Use in ComfyUI

1. Put the .safetensors file in ComfyUI/models/diffusion_models/.
2. Use the stock Krea 2 text encoder and VAE (from Comfy-Org/Krea-2) and the built-in Krea 2 text-to-image template, swapping only the diffusion model.
3. Sampling: 8 steps, CFG off (Turbo distillation).
4. Layout prompts use the grounding DSL from the source repo (included here as PROMPTING.md): one element per line, boxes as [x0,y0,x1,y1] on a 0–1000 grid, x first. Plain prose prompts also work.

Quantization details

• Tool: comfyui-native-quantizer (comfyquant), version and commit recorded below after the run.
• Format: int8-convrot (INT8 weights, FP32 row scales, ConvRot rotation), target runtime NVIDIA, conversion validated with --validate and --verify-output.
• Recipe: the layer selection and per-layer ConvRot group sizes (256/64/16) were extracted from the official Comfy-Org/Krea-2 artifact diffusion_models/krea2_turbo_int8_convrot.safetensors (revision 1171f4dc7e325bf3734586942e737872506c73e1) with the quantizer's read_comfy_org_recipe.py and applied via --reference-recipe, so these files select exactly the layers the official Krea 2 Turbo INT8 build selects.
• Source revision: jimmycarter/krea2-turbo-bbox @ 95eae22321f4aed6e499d7f7472a97cc5a8c4c20; source SHA256 verified before conversion.
• Smoke test: each checkpoint was loaded in ComfyUI and generated the grounded four-panel prompt from PROMPTING.md before upload. Sample images appended after the run.

Credits

• Base model: Krea (Krea 2 Raw / Turbo), under the Krea 2 Community License Agreement.
• BBOX layout finetune: jimmycarter.
• Quantization: NidAll, with comfyui-native-quantizer (Apache-2.0; the tool's license does not change the weights' license above).

Conversion record

• Date (UTC): 2026-11-01 14:40
• Tool: comfyquant comfyui-native-quantizer 2.3.0rc3 (comfyui-native-quantizer)
• Recipe source: Comfy-Org/Krea-2 @ 1171f4dc7e325bf3734586942e737872506c73e1, diffusion_models/krea2_turbo_int8_convrot.safetensors (224 layers)
• Source: jimmycarter/krea2-turbo-bbox @ 95eae22321f4aed6e499d7f7472a97cc5a8c4c20, SHA256 of all three inputs verified before conversion
• krea2-bbox-turbo-int8-convrot.safetensors: 14,135,120,528 bytes (official size 13,492,686,496, delta +642,434,032 bytes)
• krea2-bbox-turbo-ema_0.9995-int8-convrot.safetensors: 14,135,120,552 bytes (official size 13,492,686,496, delta +642,434,056 bytes)
• krea2-bbox-turbo-ema_0.99975-int8-convrot.safetensors: 14,135,120,568 bytes (official size 13,492,686,496, delta +642,434,072 bytes)

Output SHA256:

3cd0c03554a59816540fea208c564fe2bf9420125f53e2914b2b6135957167ae  krea2-bbox-turbo-ema_0.9995-int8-convrot.safetensors
a5b76a1d013fe336ff68a1f3a4740c63d6cca41995913382a7b89776cff0c1ba  krea2-bbox-turbo-ema_0.99975-int8-convrot.safetensors
712bde1abde232ee5f73485324eabd139ac6ba1568b07132ccaa4aaf5cfe4882  krea2-bbox-turbo-int8-convrot.safetensors