--- license: other license_name: mixed tags: - krea2 - moss-image - quantized - safetensors --- # Model cache Model files for image-model LoRA training: diffusion transformers pre-quantised to INT8 and NF4, VAEs, text encoders and a LoRA. Each file keeps the licence of the model it comes from (see [Licences](#licences)). `manifest.json` lists every file with an alias and its sha256. ## Files | Path | Alias | Content | |---|---|---| | `dit/krea2-raw.int8.k2q.safetensors` | `krea2-int8` | Krea 2 Raw DiT, block Linears quantised to INT8 (13.5 GB) | | `dit/krea2-raw.nf4.k2q.safetensors` | `krea2-nf4` | Krea 2 Raw DiT, block Linears quantised to NF4 (8.2 GB) | | `dit/moss-image-base.int8.k2q.safetensors` | `moss-image-int8` | MOSS Image Base DiT, block Linears quantised to INT8 (5.4 GB) | | `dit/moss-image-base.nf4.k2q.safetensors` | `moss-image-nf4` | MOSS Image Base DiT, block Linears quantised to NF4 (3.3 GB) | | `vae/qwen-image-vae.fp32.safetensors` | `qwen-image-vae` | Qwen-Image VAE, fp32 (Krea 2) | | `vae/qwen-image-vae.bf16.safetensors` | `qwen-image-vae-bf16` | Qwen-Image VAE, bf16 (Krea 2) | | `vae/flux2-vae.fp32.safetensors` | `flux2-vae` | FLUX.2 VAE, fp32 (MOSS Image) | | `text_encoder/qwen3vl-4b.bf16.safetensors` | `qwen3vl-4b-bf16` | Qwen3-VL-4B text encoder, bf16 (Krea 2, 8.9 GB) | | `text_encoder/qwen3vl-4b.fp8_scaled.safetensors` | `qwen3vl-4b-fp8` | Qwen3-VL-4B text encoder, fp8 scaled (Krea 2, 5.2 GB) | | `text_encoder/qwen3.5-9b.int8.safetensors` | `qwen35-9b-int8` | Qwen3.5-9B text encoder, INT8 (MOSS Image, 9.0 GB) | | `lora/krea2-turbo-r64.safetensors` | `krea2-turbo-r64` | Krea 2 turbo LoRA, rank 64, bf16: 8 steps at strength 1.0, or 12 at 0.6 | | `face/buffalo_l/` | `buffalo_l` | InsightFace face detection (`det_10g.onnx`) and recognition (`w600k_r50.onnx`) | ## Formats **Pre-quantised DiTs (`.k2q.safetensors`).** The Linear weights of the transformer blocks are quantised: - INT8: an int8 weight with a bf16 scale per output channel; - NF4: the packed 4-bit weight with its bitsandbytes `QuantState` tensors. Every other tensor is bf16, as in the source. The safetensors header metadata records the format version, the quantisation, the architecture and its config, the source file name and sha256, and a checksum of the tensor data. In the MOSS Image files each block's q, k and v are separate tensors (`layers.N.proj.q`, `.k`, `.v`), where the source fuses them into one `qkv` weight. **INT8 text encoder.** `qwen3.5-9b.int8.safetensors` is in ComfyUI's mixed-precision layout: each quantised weight is int8 with an fp32 `weight_scale` per output channel (`[out, 1]`) and a `comfy_quant` entry with format `int8_tensorwise`. Only the 2-D weights of the decoder layers are quantised. Embeddings, norms and the Gated DeltaNet's `conv1d`, `A_log`, `dt_bias`, `in_proj_a` and `in_proj_b` stay bf16. The vision tower, the MTP head and `lm_head` are removed, so it encodes text only. On four prompts its last hidden state had a cosine similarity of 0.9988 to 0.9995 with the bf16 model's. ## Sources | Files | Source | |---|---| | `dit/krea2-raw.*` | `raw.safetensors` of [krea/Krea-2-Raw](https://huggingface.co/krea/Krea-2-Raw), sha256 `f99bb0ff8e362b77342bc4994e0c50906fe7ef7074864b181b7d48d2fa6d03d7`, quantised | | `dit/moss-image-base.*` | `diffusion_models/moss_image_base_bf16.safetensors` of [Comfy-Org/MOSS-Image](https://huggingface.co/Comfy-Org/MOSS-Image) (a repack of [OpenMOSS-Team/MOSS-Image-Base](https://huggingface.co/OpenMOSS-Team/MOSS-Image-Base)), sha256 `cacd158728a1fb9e57c408a922b54752ec71fceb566b3821269927964e62b41f`, quantised | | `vae/qwen-image-vae.fp32.safetensors`, `text_encoder/qwen3vl-4b.bf16.safetensors` | `vae/` and `text_encoder/` of krea/Krea-2-Raw, unmodified | | `vae/qwen-image-vae.bf16.safetensors`, `text_encoder/qwen3vl-4b.fp8_scaled.safetensors` | ComfyUI repacks, unmodified | | `vae/flux2-vae.fp32.safetensors` | `vae/flux2-vae.safetensors` of Comfy-Org/MOSS-Image, unmodified | | `text_encoder/qwen3.5-9b.int8.safetensors` | `text_encoders/qwen3.5_9b_bf16.safetensors` of [Comfy-Org/Qwen3.5](https://huggingface.co/Comfy-Org/Qwen3.5), sha256 `7e6e9f08d598f829cb940e60ac0c698e1f1c27a47daffd7e598cd78c78b4cc53`, quantised | | `lora/krea2-turbo-r64.safetensors` | `krea2_turbo_lora_rank_64_bf16.safetensors`, unmodified | | `face/buffalo_l/` | [InsightFace](https://github.com/deepinsight/insightface) `buffalo_l` | ## Licences The licence texts and the attribution notice are in [`licenses/`](licenses): [Krea 2 Community License](licenses/Krea-2-Community-License.pdf), [Apache License 2.0](licenses/Apache-2.0.txt), [NOTICE](licenses/NOTICE). | Files | Licence | |---|---| | `dit/krea2-raw.*`, `lora/krea2-turbo-r64.safetensors` | Krea 2 Community License. By downloading or using these files you agree to be bound by it. They are modified or redistributed Krea 2 files, not official Krea products (see [NOTICE](licenses/NOTICE)). | | `dit/moss-image-base.*` | Apache License 2.0 ([MOSS Image](https://huggingface.co/OpenMOSS-Team/MOSS-Image-Base)) | | `text_encoder/qwen3.5-9b.int8.safetensors` | Apache License 2.0 ([Qwen3.5-9B](https://huggingface.co/Qwen/Qwen3.5-9B)) | | `text_encoder/qwen3vl-4b.*` | Apache License 2.0 ([Qwen3-VL-4B-Instruct](https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct)) | | `vae/qwen-image-vae.*` | Apache License 2.0 ([Qwen-Image](https://huggingface.co/Qwen/Qwen-Image)) | | `vae/flux2-vae.fp32.safetensors` | Apache License 2.0 (the FLUX.2 VAE by Black Forest Labs, published under it with [FLUX.2-klein-4B](https://huggingface.co/black-forest-labs/FLUX.2-klein-4B)) | | `face/buffalo_l/` | InsightFace pretrained models: non-commercial research use only | Changed files replace the old ones in place; the repository's revisions are the versions.