Draw Things Mirror: Shared Components
Shared components used by the JoyFusion Draw Things model catalog. Files retain their original Draw Things filenames and are mirrored byte-for-byte from static.libnnc.org.
| File | Role | Bytes | License family |
|---|---|---|---|
qwen_3_vl_4b_instruct_q8p.ckpt |
Z-Image text encoder | 4,535,328,768 | Apache-2.0 |
ltx_2.3_audio_video_vae_f16.ckpt |
LTX audio/video VAE | 1,820,950,528 | LTX-2 Community License |
ltx_2.3_spatial_upscaler_x1.5_f16.ckpt |
LTX spatial upscaler | 1,092,734,976 | LTX-2 Community License |
ltx_2.3_spatial_upscaler_x2_1.1_f16.ckpt |
LTX spatial upscaler | 998,080,512 | LTX-2 Community License |
flux_1_vae_f16.ckpt |
Z-Image/FLUX-compatible VAE | 167,870,464 | See per-file notices |
qwen_3_0.6b_instruct_q8p.ckpt |
Qwen3-0.6B (Instruct) text encoder, 8-bit β JoyFusion conversion, see below | 767,348,736 | Apache-2.0 |
Because this repository contains components governed by different terms, it intentionally uses license: other. Applicable upstream license texts and notices are provided under LICENSES/; users must comply with the terms for each file they use.
This repository contains converted/quantized artifacts for Draw Things compatibility. It is not affiliated with the original model authors or Draw Things.
JoyFusion conversions
qwen_3_0.6b_instruct_q8p.ckpt is not a Draw Things file. JoyFusion converted Qwen/Qwen3-0.6B (revision c1899de2, Apache-2.0) to the Draw Things text-encoder format, using the official qwen_3_0.6b_q8p.ckpt (Qwen3-0.6B-Base) as the template: tensor names, shapes and codecs are identical (linear layers 8-bit palettized, norms ezm7, token embedding fp16), and the q/k projections are reordered for Draw Things' interleaved rotary layout. Every tensor was checked against the source weights (largest relative error 0.67%, from the 8-bit quantization). It is the text encoder for the Qwen3-0.6B β FLUX.2 [klein] 4B adapter in JoyFusionAI/dt-flux2-klein-4b.
SHA-256: 2e71a423b7e1e774585095f14563ca727d489983c4b044db71a2bb93db2e56aa