This is gamechanging. Wow.
After watching experiments on Textual Inversion from the sidelines, I decided to jump in and try this. Thank goodness I had a A6000 on hand for the whopping 38 gigs of VRAM needed. but the results are amazing. Far better than basic Textual Inversion. And it trains in a FRACTION of the time! the following images are me as a zombie, using a basic prompt like “digital painting of yourNameHere as a zombie, undead, decomposing, concept art, by Greg Rutkowski and Wlop”    some of my initial findings: Low CFG (I used 5) helps the image be more flexible in style and subject. Otherwise it kind of just looks like the iPhone selfies I trained it on. You really have to push to bring in painter styles and subjects. I had to write “zombie, undead, decomposing” before it finally broke away from generating a normal face. there are about a million variables I can think of to start tweaking. I would love to hear about other people’s experiences. What’s the best image set to train on? How many images? How many iterations? Can we make this run on a 3090? I have a t
Posted by nikopueringer on Sep 8, 2022, 11:01 PM
Tags
Related issues
- struct's lifetime requirements can be bypassed via covariant field assignment
- Proposal to Reorganize the Moderation Team and Review Moderation Policies
- Progressive FPS degradation appears to be related to dota2.com — blocking the domain prevents the performance loss
- proposal: container/...: generic collection types
- Better Custom Endpoint BYOK
- A concrete policy for backward compatibility until Godot 5.0
Raise real-world problems, find people facing the same issue, and discuss solutions.2,500 issues0 comments
github (Other)
struct's lifetime requirements can be bypassed via covariant field assignment
Proposal to Reorganize the Moderation Team and Review Moderation Policies
Progressive FPS degradation appears to be related to dota2.com — blocking the domain prevents the performance loss
proposal: container/...: generic collection types
Better Custom Endpoint BYOK
A concrete policy for backward compatibility until Godot 5.0
New MTP models not supported
Tracking: 429 / Capacity Issues
This is gamechanging. Wow.
After watching experiments on Textual Inversion from the sidelines, I decided to jump in and try this. Thank goodness I had a A6000 on hand for the whopping 38 gigs of VRAM needed. but the results are amazing. Far better than basic Textual Inversion. And it trains in a FRACTION of the time! the following images are me as a zombie, using a basic prompt like “digital painting of yourNameHere as a zombie, undead, decomposing, concept art, by Greg Rutkowski and Wlop”    some of my initial findings: Low CFG (I used 5) helps the image be more flexible in style and subject. Otherwise it kind of just looks like the iPhone selfies I trained it on. You really have to push to bring in painter styles and subjects. I had to write “zombie, undead, decomposing” before it finally broke away from generating a normal face. there are about a million variables I can think of to start tweaking. I would love to hear about other people’s experiences. What’s the best image set to train on? How many images? How many iterations? Can we make this run on a 3090? I have a t
Comments (0)
No comments yet. Start the discussion.