Hugging Face's logo

RichardErkhov
/
KBlueLeaf_-_TIPO-200M-ft-gguf

GGUF
Inference Endpoints
Model card Files Community
RichardErkhov commited on
Commit
32b299e
·
verified ·
1 Parent(s): 33a41a3

uploaded readme

Browse files
Files changed (1) hide show
  1. README.md +147 -0
README.md ADDED
@@ -0,0 +1,147 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Quantization made by Richard Erkhov.
2
+
3
+ [Github](https://github.com/RichardErkhov)
4
+
5
+ [Discord](https://discord.gg/pvy7H8DZMG)
6
+
7
+ [Request more models](https://github.com/RichardErkhov/quant_request)
8
+
9
+
10
+ TIPO-200M-ft - GGUF
11
+ - Model creator: https://huggingface.co/KBlueLeaf/
12
+ - Original model: https://huggingface.co/KBlueLeaf/TIPO-200M-ft/
13
+
14
+
15
+ | Name | Quant method | Size |
16
+ | ---- | ---- | ---- |
17
+ | [TIPO-200M-ft.Q2_K.gguf](https://huggingface.co/RichardErkhov/KBlueLeaf_-_TIPO-200M-ft-gguf/blob/main/TIPO-200M-ft.Q2_K.gguf) | Q2_K | 0.08GB |
18
+ | [TIPO-200M-ft.IQ3_XS.gguf](https://huggingface.co/RichardErkhov/KBlueLeaf_-_TIPO-200M-ft-gguf/blob/main/TIPO-200M-ft.IQ3_XS.gguf) | IQ3_XS | 0.09GB |
19
+ | [TIPO-200M-ft.IQ3_S.gguf](https://huggingface.co/RichardErkhov/KBlueLeaf_-_TIPO-200M-ft-gguf/blob/main/TIPO-200M-ft.IQ3_S.gguf) | IQ3_S | 0.09GB |
20
+ | [TIPO-200M-ft.Q3_K_S.gguf](https://huggingface.co/RichardErkhov/KBlueLeaf_-_TIPO-200M-ft-gguf/blob/main/TIPO-200M-ft.Q3_K_S.gguf) | Q3_K_S | 0.09GB |
21
+ | [TIPO-200M-ft.IQ3_M.gguf](https://huggingface.co/RichardErkhov/KBlueLeaf_-_TIPO-200M-ft-gguf/blob/main/TIPO-200M-ft.IQ3_M.gguf) | IQ3_M | 0.09GB |
22
+ | [TIPO-200M-ft.Q3_K.gguf](https://huggingface.co/RichardErkhov/KBlueLeaf_-_TIPO-200M-ft-gguf/blob/main/TIPO-200M-ft.Q3_K.gguf) | Q3_K | 0.1GB |
23
+ | [TIPO-200M-ft.Q3_K_M.gguf](https://huggingface.co/RichardErkhov/KBlueLeaf_-_TIPO-200M-ft-gguf/blob/main/TIPO-200M-ft.Q3_K_M.gguf) | Q3_K_M | 0.1GB |
24
+ | [TIPO-200M-ft.Q3_K_L.gguf](https://huggingface.co/RichardErkhov/KBlueLeaf_-_TIPO-200M-ft-gguf/blob/main/TIPO-200M-ft.Q3_K_L.gguf) | Q3_K_L | 0.1GB |
25
+ | [TIPO-200M-ft.IQ4_XS.gguf](https://huggingface.co/RichardErkhov/KBlueLeaf_-_TIPO-200M-ft-gguf/blob/main/TIPO-200M-ft.IQ4_XS.gguf) | IQ4_XS | 0.11GB |
26
+ | [TIPO-200M-ft.Q4_0.gguf](https://huggingface.co/RichardErkhov/KBlueLeaf_-_TIPO-200M-ft-gguf/blob/main/TIPO-200M-ft.Q4_0.gguf) | Q4_0 | 0.11GB |
27
+ | [TIPO-200M-ft.IQ4_NL.gguf](https://huggingface.co/RichardErkhov/KBlueLeaf_-_TIPO-200M-ft-gguf/blob/main/TIPO-200M-ft.IQ4_NL.gguf) | IQ4_NL | 0.11GB |
28
+ | [TIPO-200M-ft.Q4_K_S.gguf](https://huggingface.co/RichardErkhov/KBlueLeaf_-_TIPO-200M-ft-gguf/blob/main/TIPO-200M-ft.Q4_K_S.gguf) | Q4_K_S | 0.11GB |
29
+ | [TIPO-200M-ft.Q4_K.gguf](https://huggingface.co/RichardErkhov/KBlueLeaf_-_TIPO-200M-ft-gguf/blob/main/TIPO-200M-ft.Q4_K.gguf) | Q4_K | 0.12GB |
30
+ | [TIPO-200M-ft.Q4_K_M.gguf](https://huggingface.co/RichardErkhov/KBlueLeaf_-_TIPO-200M-ft-gguf/blob/main/TIPO-200M-ft.Q4_K_M.gguf) | Q4_K_M | 0.12GB |
31
+ | [TIPO-200M-ft.Q4_1.gguf](https://huggingface.co/RichardErkhov/KBlueLeaf_-_TIPO-200M-ft-gguf/blob/main/TIPO-200M-ft.Q4_1.gguf) | Q4_1 | 0.12GB |
32
+ | [TIPO-200M-ft.Q5_0.gguf](https://huggingface.co/RichardErkhov/KBlueLeaf_-_TIPO-200M-ft-gguf/blob/main/TIPO-200M-ft.Q5_0.gguf) | Q5_0 | 0.13GB |
33
+ | [TIPO-200M-ft.Q5_K_S.gguf](https://huggingface.co/RichardErkhov/KBlueLeaf_-_TIPO-200M-ft-gguf/blob/main/TIPO-200M-ft.Q5_K_S.gguf) | Q5_K_S | 0.13GB |
34
+ | [TIPO-200M-ft.Q5_K.gguf](https://huggingface.co/RichardErkhov/KBlueLeaf_-_TIPO-200M-ft-gguf/blob/main/TIPO-200M-ft.Q5_K.gguf) | Q5_K | 0.14GB |
35
+ | [TIPO-200M-ft.Q5_K_M.gguf](https://huggingface.co/RichardErkhov/KBlueLeaf_-_TIPO-200M-ft-gguf/blob/main/TIPO-200M-ft.Q5_K_M.gguf) | Q5_K_M | 0.14GB |
36
+ | [TIPO-200M-ft.Q5_1.gguf](https://huggingface.co/RichardErkhov/KBlueLeaf_-_TIPO-200M-ft-gguf/blob/main/TIPO-200M-ft.Q5_1.gguf) | Q5_1 | 0.14GB |
37
+ | [TIPO-200M-ft.Q6_K.gguf](https://huggingface.co/RichardErkhov/KBlueLeaf_-_TIPO-200M-ft-gguf/blob/main/TIPO-200M-ft.Q6_K.gguf) | Q6_K | 0.16GB |
38
+ | [TIPO-200M-ft.Q8_0.gguf](https://huggingface.co/RichardErkhov/KBlueLeaf_-_TIPO-200M-ft-gguf/blob/main/TIPO-200M-ft.Q8_0.gguf) | Q8_0 | 0.2GB |
39
+
40
+
41
+
42
+
43
+ Original model description:
44
+ ---
45
+ license: other
46
+ license_name: kohaku-license-1.0
47
+ datasets:
48
+ - laion/conceptual-captions-12m-webdataset
49
+ - CaptionEmporium/coyo-hd-11m-llavanext
50
+ - KBlueLeaf/danbooru2023-metadata-database
51
+ - graph-based-captions/GBC10M
52
+ language:
53
+ - en
54
+ pipeline_tag: text-generation
55
+ library_name: transformers
56
+ ---
57
+ # TIPO: Text to Image with text presampling for Prompt Optimization
58
+
59
+ 200M LLaMA arch model trained for TIPO. <br>
60
+ Tech Report: https://arxiv.org/abs/2411.08127
61
+
62
+ ![image/png](https://cdn-uploads.huggingface.co/production/uploads/630593e2fca1d8d92b81d2a1/fc9ovmARapQmgq9DZ7ApJ.png)
63
+
64
+ ## Introduction
65
+
66
+ In this project, we introduce "TIPO" (**T**ext to **I**mage with text presampling for **P**rompt **O**ptimization), an innovative framework designed to significantly enhance the quality and usability of Text-to-Image (T2I) generative models. TIPO utilizes the Large Language Models (LLMs) to perform "Text Presampling" within the inference pipeline of text-to-image generative modeling. By refining and extending user input prompts, TIPO enables generative models to produce superior results with minimal user effort, making T2I systems more accessible and effective for a wider range of users.
67
+
68
+ ## Usage
69
+
70
+ Use updated version of DTG extension (renamed to z-tipo-extension), current version of z-tipo-extension support stable-diffusion-webui, stable-diffusion-webui-forge and ComfyUI. SD-Next haven't been tested.
71
+ https://github.com/KohakuBlueleaf/z-tipo-extension
72
+
73
+ ## Model arch and Training
74
+
75
+ This model is LLaMA arch with 200M parameters, the training data is combined version of Danbooru2023, Coyo-HD-11M. <br>
76
+ The total token seen is around 50B tokens. <br>
77
+ For more information please refer to the tech report and following table.
78
+
79
+ | | TIPO-200M | TIPO-200M-ft | TIPO-500M |
80
+ | ----------------- | ------------------------------------------------------------------------------ | ---------------------------------- | ------------------------------------------------------------------------------ |
81
+ | Arch | LLaMA | LLaMA | LLaMA |
82
+ | Max ctx length | 1024 | 1024 | 1024 |
83
+ | Batch Size | 2048 | 2048 | 3584 |
84
+ | Training dataset | Danbooru, GBC10M, 5epoch<br />Danbooru, GBC10M, Coyo11M, 3epoch | Danbooru(pixtral), Coyo11M, 2epoch | Danbooru, GBC10M, Coyo11M, 5epoch |
85
+ | Real Token Seen* | 40B token | 50B (10B more from TIPO-200M) | 30B token |
86
+ | Training Hardware | RTX 3090 x 4 | RTX 3090 x 4 | H100 x 8 |
87
+ | Training Time | 420 hour` | 120 hour` | 100 hour` |
88
+ | Huggingface | [KBlueLeaf/TIPO-200M · Hugging Face](https://huggingface.co/KBlueLeaf/TIPO-200M) | You Are HERE | [KBlueLeaf/TIPO-500M · Hugging Face](https://huggingface.co/KBlueLeaf/TIPO-500M) |
89
+
90
+ *: We only count "non-padding token" in the token seen, since all the training data have very large length range. <br>
91
+ `: Since the training data is pretty short, it cost more time to reach same token seen than general LLM pretraining. <br>
92
+ As reference, with 4096 as max ctx length and almost all the data have reach that length, you may only need 2days to reach 10B token seen on RTX 3090 x 4 with 200M model.
93
+
94
+ ### Evaluation
95
+ **Evaluation are done on TIPO-200M model** <br>
96
+ We have tested TIPO compared to other Model in several test and metrics:
97
+
98
+ #### Scenery tag test
99
+
100
+ In this test we use single "scenery" tag as input. (With some certain meta) <br>
101
+ To test each prompt gen method to see if they can obtain the desired distribution of outputs while maintain the quality of images.
102
+
103
+ | Scenery Tag Test | Original | GPT4o-mini | Prompt DB | Promptis | TIPO(ours) |
104
+ | ---- | ---- | ---- | ---- | ---- | ---- |
105
+ | FDD ↓ | 0.3558 | 0.5414 | 0.3247 | *0.2350* | **0.2282** |
106
+ | Aesthetic ↑ | 5.0569 | **6.3676** | 6.1609 | 5.9468 | *6.2571* |
107
+ | AI Corrupt ↑ | 0.4257 | *0.7490* | 0.5024 | 0.5669 | **0.9195** |
108
+
109
+ #### Short/Truncated Long test
110
+
111
+ In this test we use short caption or manually truncated caption from GBC10M and CoyoHD11M. <br>
112
+ This test examine the ability of prompt gen method on handling almostly completed prompts.
113
+
114
+ | Short | Original | GPT4o-mini | Prompt DB | Promptis | TIPO(ours) |
115
+ | ---- | ---- | ---- | ---- | ---- | ---- |
116
+ | FDD ↓ | 0.0957 | 0.1668 | *0.0980* | 0.1783 | 0.1168 |
117
+ | Aesthetic ↑ | 5.8370 | **6.0589** | 5.8213 | 5.7963 | *5.8531* |
118
+ | AI Corrupt ↑ | 0.7113 | 0.6985 | 0.7064 | 0.6314 | **0.7131** |
119
+
120
+ | Truncated Long | Original | GPT4o-mini | Prompt DB | Promptis | TIPO(ours) |
121
+ | ---- | ---- | ---- | ---- | ---- | ---- |
122
+ | FDD ↓ | 0.0955 | 0.1683 | *0.1247* | 0.2096 | 0.1210 |
123
+ | Aesthetic ↑ | 5.7497 | **6.0168** | 5.8191 | 5.7759 | *5.8364* |
124
+ | AI Corrupt ↑ | 0.6868 | 0.6712 | 0.6741 | 0.5925 | **0.7130** |
125
+
126
+
127
+
128
+ ## LICENSE
129
+
130
+ This model is released under [Kohaku License 1.0](https://kblueleaf.net/documents/kohaku-license/?[Your%20Organization/Name]=KohakuBlueLeaf&[Year]=2024) <br>
131
+ You can check the above provided URL or check the LICENSE file in this repo.
132
+
133
+ ### Citation
134
+
135
+ ```bibtex
136
+ @misc{yeh2024tipotextimagetext,
137
+ title={TIPO: Text to Image with Text Presampling for Prompt Optimization},
138
+ author={Shih-Ying Yeh and Sang-Hyun Park and Giyeong Oh and Min Song and Youngjae Yu},
139
+ year={2024},
140
+ eprint={2411.08127},
141
+ archivePrefix={arXiv},
142
+ primaryClass={cs.CV},
143
+ url={https://arxiv.org/abs/2411.08127},
144
+ }
145
+ ```
146
+
147
+