devon7y commited on
Commit
be29bfc
·
verified ·
1 Parent(s): 89c2798

WikiQwen-Illustrator-9B: final round-2 weights (step 18000)

Browse files
Files changed (2) hide show
  1. README.md +3 -4
  2. pytorch_lora_weights.safetensors +1 -1
README.md CHANGED
@@ -23,8 +23,6 @@ tags:
23
 
24
  # WikiQwen-Illustrator-9B
25
 
26
- > **Preview weights.** Round 2 at step 6,000 of 18,000 (full ~588k-image set, rank 128). Training continues and the final weights will replace these.
27
-
28
  **Give it a caption, get a step illustration.** WikiQwen-Illustrator-9B is the larger sibling of [WikiQwen-Illustrator](https://huggingface.co/devon7y/WikiQwen-Illustrator): a style LoRA for [FLUX.2 [klein] base 9B](https://huggingface.co/black-forest-labs/FLUX.2-klein-base-9B). It turns a one-to-three-sentence scene description into a clean, friendly how-to illustration at 512x384. It is the picture half of WikiQwen: the chat models ([4B](https://huggingface.co/devon7y/WikiQwen-4B), [9B](https://huggingface.co/devon7y/WikiQwen-9B), [35B-A3B](https://huggingface.co/devon7y/WikiQwen-35B-A3B)) write an `[IMAGE: caption]` line above every step, and this LoRA draws it.
29
 
30
  > **Try it first:** [WikiQwen Plus](https://huggingface.co/spaces/devon7y/wikiqwen-plus) (WikiQwen-4B) and [WikiQwen Pro](https://huggingface.co/spaces/devon7y/wikiqwen-pro) (WikiQwen-9B) write an article, then draw every step with this LoRA.
@@ -116,8 +114,9 @@ Each held-out caption was drawn once and compared with the real illustration for
116
 
117
  | Setup | Images | CMMD ↓ | KID ×10³ ↓ | Real-vs-generated probe ↓ | Caption alignment ↑ (generated / real) |
118
  | --- | ---: | ---: | ---: | ---: | ---: |
119
- | Preview: 9B round 2 step 6000 | 300 | 0.182 | 40.52 ± 36.70 | 0.963 ± 0.007 | 0.2097 / 0.1965 |
120
- | 9B round 1 bake-off (step 3000) | 300 | 0.175 | 91.40 ± 42.18 | 0.972 ± 0.015 | 0.2109 / 0.1965 |
 
121
  | WikiQwen-Illustrator 4B (round 2 final) | 300 | 0.153 | 70.85 ± 39.05 | 0.970 ± 0.015 | 0.2053 / 0.1965 |
122
  | Real vs. real (two halves of the held-out set) | 150 + 150 | 0.041 | -23.53 | 0.5 = chance | 0.1965 (real) |
123
 
 
23
 
24
  # WikiQwen-Illustrator-9B
25
 
 
 
26
  **Give it a caption, get a step illustration.** WikiQwen-Illustrator-9B is the larger sibling of [WikiQwen-Illustrator](https://huggingface.co/devon7y/WikiQwen-Illustrator): a style LoRA for [FLUX.2 [klein] base 9B](https://huggingface.co/black-forest-labs/FLUX.2-klein-base-9B). It turns a one-to-three-sentence scene description into a clean, friendly how-to illustration at 512x384. It is the picture half of WikiQwen: the chat models ([4B](https://huggingface.co/devon7y/WikiQwen-4B), [9B](https://huggingface.co/devon7y/WikiQwen-9B), [35B-A3B](https://huggingface.co/devon7y/WikiQwen-35B-A3B)) write an `[IMAGE: caption]` line above every step, and this LoRA draws it.
27
 
28
  > **Try it first:** [WikiQwen Plus](https://huggingface.co/spaces/devon7y/wikiqwen-plus) (WikiQwen-4B) and [WikiQwen Pro](https://huggingface.co/spaces/devon7y/wikiqwen-pro) (WikiQwen-9B) write an article, then draw every step with this LoRA.
 
114
 
115
  | Setup | Images | CMMD ↓ | KID ×10³ ↓ | Real-vs-generated probe ↓ | Caption alignment ↑ (generated / real) |
116
  | --- | ---: | ---: | ---: | ---: | ---: |
117
+ | This release: 9B round 2 final (step 18000) | 300 | 0.177 | 33.67 ± 37.13 | 0.968 ± 0.012 | 0.2093 / 0.1965 |
118
+ | Earlier preview: 9B round 2 step 6000 | 300 | 0.182 | 40.52 ± 36.70 | 0.963 ± 0.007 | 0.2097 / 0.1965 |
119
+ | This release drawn at 768x576 (scored after downscaling to 512x384) | 300 | 0.157 | 38.53 ± 36.34 | 0.967 ± 0.014 | 0.2108 / 0.1965 |
120
  | WikiQwen-Illustrator 4B (round 2 final) | 300 | 0.153 | 70.85 ± 39.05 | 0.970 ± 0.015 | 0.2053 / 0.1965 |
121
  | Real vs. real (two halves of the held-out set) | 150 + 150 | 0.041 | -23.53 | 0.5 = chance | 0.1965 (real) |
122
 
pytorch_lora_weights.safetensors CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:bd05da0e098c29bdc4f451a0f43bc02f27aa1df8376ed480ed2bed225bd1c0d0
3
  size 1392551856
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b5bd8e878478001bbd7c49a5d4e79502ac90b2f926c5d77564b9b1b7e28456a5
3
  size 1392551856