This is a result from a model that combines LCM, SDXL, custom loras and multi-controlnet, ~4x faster. 1024×1024, 8 steps, HED controlnet: 3.86s
@fofrai
-
LCM Model Distillation Quality Trade-offs with LoRA
By
–
The image quality of the fully distilled model (ie the original LCM) is really good, though it kind of forces a particular style. With the lora loaded versions, the quality takes a hit. I'm waiting for more fully distilled models to be released by @SimianLuo
-
Latent Consistency Model 3x Faster on Replicate
By
–
I've optimised my latent consistency model on Replicate to make it 3x faster: https://
replicate.com/fofr/latent-co
nsistency-model
… text to image, 4 steps at 768×768:
– was: 2s
– now: 0.6s -
AI Fine-Tune Creates Stunning 70s-Inspired Thanksgiving Food Photography
By
–
Do you need some thanksgiving meal inspiration? @CarlPeaslee convinced me to make a fine-tune based on some beautiful 70s food photography. I decided to give this a go on @everartai
, and the results do not disappoint. -
GitHub Copilot Shell Alias for Quick Command Suggestions
By
–
I liked the `??` way of doing things, so I made an alias: echo "alias ??='gh copilot suggest -t shell'" >> ~/.bashrc So I can now write this and get to the answer in one hop: `?? disk space`
-
Regional Prompting and Attention Mechanisms in AI Models
By
–
I wonder if this is based on some variant of regional prompting 🤔
— fofr (@fofrAI) 17 novembre 2023
(The attention mechanism that underlies it) https://t.co/yDP1EfraLGI wonder if this is based on some variant of regional prompting (The attention mechanism that underlies it)
-
Attention Mechanism in AI: Technical Implementation
By
–
I mean the attention mechanism to achieve it.
-
SDXL+LCM Guidance Parameter Optimization Experiments
By
–
Running experiments to see where the guidance sweet spot lies for SDXL+LCM lora.
— fofr (@fofrAI) 16 novembre 2023
Here's a video running from 0 to 3 with 0.01 increments.
No change between 0 and 1, then the astronaut goes wild. pic.twitter.com/icGnaaEBTnRunning experiments to see where the guidance sweet spot lies for SDXL+LCM lora. Here's a video running from 0 to 3 with 0.01 increments. No change between 0 and 1, then the astronaut goes wild.
-
API Base64 Image Processing: Beyond the Bottleneck
By
–
The API accepts base64 images, but I don't think that's the bottleneck here.
-
GitHub Copilot CLI Public Beta Criticized for Reduced Functionality
By
–
I'm not keen on the new Github Copilot CLI release. More hoops to jump through to do the same thing? Left is the public beta – it does not explain the command, or give an option to run it. Copy command fails on a VM Right is the earlier closed beta – automatic explanation,