This is a raw comparison of 5 different Q8 GGUF models using the exact same seed and the exact same prompt for each test. The settings are exactly the same for each test run: 8 steps, 1.0 cfg, euler, normal. All of these are 12.7 GB. I will do several different prompts in the same exact format over time once I find a prompt I think will work well, and then compare them all at the end. I also added the base Flux.1 D model. My settings for Dev were: 20 steps, 1.0 cfg, euler, normal.
Core Hardware:
CPU: Ryzen 7 7800X3D 8-Core
GPU: RTX 4070 Ti Super Trinity OC
RAM: DDR5 32GB (2x16GB) 7200MHz
-
Prompt 1: Each sitting at a business meeting.
There are four individual characters each sitting at a business meeting. They are Pikachu, Darth Vader, Garfield, and Mickey Mouse.
-
AbdallahAlswa80_SPEED_Q8: https://civitai.com/models/661102/speedq8
Prompt executed in 65.22 seconds
-
Anibaaal_FusionDS_v0_Q8: https://civitai.com/models/630820?modelVersionId=736153
Prompt executed in 66.41 seconds
-
nakif0968_flux-schnell-dev-merge-q8-0: https://civitai.com/models/657607?modelVersionId=747918
Prompt executed in 70.45 seconds
-
RalFinger_flux1-dev-Q8_0: https://civitai.com/models/647237?modelVersionId=724149
Prompt executed in 70.27 seconds
-
RalFinger_flux1-schnell-Q8_0: https://civitai.com/models/648580?modelVersionId=725877
Prompt executed in 68.07 seconds
-
flux1-dev: https://civitai.com/models/618692/flux?modelVersionId=691639
Prompt executed in 122.80 seconds
-
Prompt 1 Direct Comparisons:
AbdallahAlswa80_SPEED_Q8
Anibaaal_FusionDS_v0_Q8
nakif0968_flux-schnell-dev-merge-q8-0
RalFinger_flux1-dev-Q8_0
RalFinger_flux1-schnell-Q8_0
flux1-dev-
Prompt 2: Ink swirl fashion model.
beautiful attractive female model, she is wearing an avante garde weird vibrant colorful ink swirl baggy pants and a curtain crop top, her face is blank covered by an opaque white box, wearing tall high heel colorful plastic boots, her hands are up above her head, whimsical bubblegum candyland background,
(white square over her face that has text says "jaykrown", signed word jaykrown),
very high quality, high detail, high resolution, ultra high definition, best quality, 4k, deep shadows, HDR, deep shadows, high dynamic range, deep colors
-
AbdallahAlswa80_SPEED_Q8: https://civitai.com/models/661102/speedq8
Prompt executed in 68.73 seconds
-
Anibaaal_FusionDS_v0_Q8: https://civitai.com/models/630820?modelVersionId=736153
Prompt executed in 60.96 seconds
-
nakif0968_flux-schnell-dev-merge-q8-0: https://civitai.com/models/657607?modelVersionId=747918
Prompt executed in 65.75 seconds
-
RalFinger_flux1-dev-Q8_0: https://civitai.com/models/647237?modelVersionId=724149
Prompt executed in 65.29 seconds
-
RalFinger_flux1-schnell-Q8_0: https://civitai.com/models/648580?modelVersionId=725877
Prompt executed in 66.96 seconds
-
flux1-dev: https://civitai.com/models/618692/flux?modelVersionId=691639
Prompt executed in 123.18 seconds
-
Prompt 2 Direct Comparisons:
AbdallahAlswa80_SPEED_Q8: Anibaaal_FusionDS_v0_Q8:nakif0968_flux-schnell-dev-merge-q8-0: RalFinger_flux1-dev-Q8_0: RalFinger_flux1-schnell-Q8_0: flux1-dev:-
Prompt 3: Text pangram.
text that says "Waltz, bad nymph, for quick jigs vex."
-
AbdallahAlswa80_SPEED_Q8: https://civitai.com/models/661102/speedq8
Prompt executed in 58.17 seconds
-
Anibaaal_FusionDS_v0_Q8: https://civitai.com/models/630820?modelVersionId=736153
Prompt executed in 69.59 seconds
-
nakif0968_flux-schnell-dev-merge-q8-0: https://civitai.com/models/657607?modelVersionId=747918
Prompt executed in 70.47 seconds
-
RalFinger_flux1-dev-Q8_0: https://civitai.com/models/647237?modelVersionId=724149
Prompt executed in 65.75 seconds
-
RalFinger_flux1-schnell-Q8_0: https://civitai.com/models/648580?modelVersionId=725877
Prompt executed in 63.93 seconds
-
flux1-dev: https://civitai.com/models/618692/flux?modelVersionId=691639
Prompt executed in 121.16 seconds
-
Prompt 3 Direct Comparisons:
AbdallahAlswa80_SPEED_Q8: Anibaaal_FusionDS_v0_Q8: nakif0968_flux-schnell-dev-merge-q8-0: RalFinger_flux1-dev-Q8_0: RalFinger_flux1-schnell-Q8_0: flux1-dev: -
Observation:
Prompt 1:
This is my subjective opinion; hopefully these images help decision-making. There are clear differences, although the same seed and same settings were used. The times were around the same with only about a 5-second difference at most. I enlarged the 1 image that I think adhered best to the prompt from the 5 total. Almost every generation adhered to the prompt, the worst performer was 'RalFinger_flux1-schnell-Q8_0'. If you look closely 'Anibaaal_FusionDS_v0_Q8' was the only model to get the prompt correctly all 5 times with each character distinctly identifiable, where all the others had duplicates or some error. The best are 'Anibaaal_FusionDS_v0_Q8' and 'AbdallahAlswa80_SPEED_Q8'. The fastest is 'AbdallahAlswa80_SPEED_Q8' at 65.22 seconds.
Prompt 2:
This was a cool test because there was a specific text requirement. "jaykrown" on a white square over her face. You can easily see which ones adhered. I don't know what went wrong with 'RalFinger_flux1-dev-Q8_0' on this one, but it performed terribly, by far the worst. The best is 'nakif0968_flux-schnell-dev-merge-q8-0' for prompt adherence. The best for style/detail are 'AbdallahAlswa80_SPEED_Q8' and 'Anibaaal_FusionDS_v0_Q8'. The fastest is 'Anibaaal_FusionDS_v0_Q8' at 60.96 seconds.
Prompt 3:
This was a test of writing ability, using a pangram "Waltz, bad nymph, for quick jigs vex." which uses one letter of the alphabet at least once. The one that got it entirely correct twice was 'AbdallahAlswa80_SPEED_Q8'. All the others struggled in other ways, some got very close but contained extra commas or duplicate words. 'nakif0968_flux-schnell-dev-merge-q8-0' and 'Anibaaal_FusionDS_v0_Q8' almost got it correct one time, each with commas after the word "quick". It doesn't seem like 'RalFinger_flux1-dev-Q8_0' and 'RalFinger_flux1-schnell-Q8_0' got it correct a single time unfortunately.
Flux.1 D:
The results speak for themselves. It's clear that the base model is still good, but two of the Q8s still hold up against it. The first test it got entirely correct I'd say. The second test it struggled with "jaykrown" on the sign, only getting it a single time, and the patterns weren't nearly as interesting as two of the Q8s. The third test it had two successes, one of which was perfect, the punctuation being the problem in the other. On average it seems like the 20 steps with the base model took almost twice as long to generate. ~120 seconds vs. ~65 seconds.
If you have any other GGUF models you think I should test let me know in the comments. I attached a .zip of all the files. I had to downscale the array images because they were each 5 MB in size.