ChrisColeTech commited on
Commit
d34024a
Β·
verified Β·
1 Parent(s): f604b9d

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +3 -3
README.md CHANGED
@@ -12,7 +12,7 @@ tags:
12
 
13
  # Mage-Flow Turbo + Edit Turbo
14
 
15
- 4-step photographic text-to-image (**~3 s** per 1024Β² image on an RTX 5090)
16
  and a companion instruction-editing model that takes up to **three**
17
  reference images.
18
 
@@ -43,7 +43,7 @@ lighting and camera are preserved; only what the instruction names changes.
43
 
44
  <table>
45
  <tr><td><img src="https://huggingface.co/ChrisColeTech/Mage-Flow/resolve/main/samples/portrait.png" width="250" alt="reference"></td><td><img src="https://huggingface.co/ChrisColeTech/Mage-Flow/resolve/main/samples/edit-hat.png" width="250" alt="edit: put a brown flat cap on his head"></td><td><img src="https://huggingface.co/ChrisColeTech/Mage-Flow/resolve/main/samples/edit-beach.png" width="250" alt="edit: change the background to a sunny beach"></td></tr>
46
- <tr><td><sub>**reference** β€” the seed-99 portrait above</sub></td><td><sub>**instruction:** `put a brown flat cap on his head` β€” 4 steps, seed 7 (~88 s)</sub></td><td><sub>**instruction:** `change the background to a sunny beach with the ocean behind him` β€” 4 steps, seed 7 (~60 s)</sub></td></tr>
47
  </table>
48
 
49
  Both edits keep the subject essentially pixel-identical β€” same skin detail,
@@ -69,7 +69,7 @@ Values these builds are run with in practice.
69
 
70
  - **Four steps is correct.** Both variants are step-distilled at guidance 1.0;
71
  raising either is not the lever for quality.
72
- - **Editing costs far more than generating** β€” ~60–90 s versus ~3 s, because
73
  the reference must be encoded through the VL tower before denoising begins.
74
  That asymmetry is inherent to the architecture, not a misconfiguration.
75
  - **Write edits as instructions, not descriptions** β€” `put a brown flat cap on
 
12
 
13
  # Mage-Flow Turbo + Edit Turbo
14
 
15
+ 4-step photographic text-to-image (**β‰ˆ3 s** per 1024Β² image on an RTX 5090)
16
  and a companion instruction-editing model that takes up to **three**
17
  reference images.
18
 
 
43
 
44
  <table>
45
  <tr><td><img src="https://huggingface.co/ChrisColeTech/Mage-Flow/resolve/main/samples/portrait.png" width="250" alt="reference"></td><td><img src="https://huggingface.co/ChrisColeTech/Mage-Flow/resolve/main/samples/edit-hat.png" width="250" alt="edit: put a brown flat cap on his head"></td><td><img src="https://huggingface.co/ChrisColeTech/Mage-Flow/resolve/main/samples/edit-beach.png" width="250" alt="edit: change the background to a sunny beach"></td></tr>
46
+ <tr><td><sub>**reference** β€” the seed-99 portrait above</sub></td><td><sub>**instruction:** `put a brown flat cap on his head` β€” 4 steps, seed 7 (β‰ˆ88 s)</sub></td><td><sub>**instruction:** `change the background to a sunny beach with the ocean behind him` β€” 4 steps, seed 7 (β‰ˆ60 s)</sub></td></tr>
47
  </table>
48
 
49
  Both edits keep the subject essentially pixel-identical β€” same skin detail,
 
69
 
70
  - **Four steps is correct.** Both variants are step-distilled at guidance 1.0;
71
  raising either is not the lever for quality.
72
+ - **Editing costs far more than generating** β€” β‰ˆ60–90 s versus β‰ˆ3 s, because
73
  the reference must be encoded through the VL tower before denoising begins.
74
  That asymmetry is inherent to the architecture, not a misconfiguration.
75
  - **Write edits as instructions, not descriptions** β€” `put a brown flat cap on