webshart/ScaleEdit-12M-metadata
Viewer • Updated • 11.8M • 52 • 1
AI & ML interests
None defined yet.
Recent Activity
View all activity
bghira
updated a
dataset 6 days ago
bghira
published a
dataset 6 days ago
Post
176
Training directly onto the Flux 3 Klein Image distilled model produces better results than finetuning the base!
I've put together a collection (https://huggingface.co/collections/SimpleTuner/flux-3-klein-image-lora-training) containing an experiment log book, a couple character LoRAs, and a 50,000 step diverse photo aesthetics finetune for the distilled variant in both 512px and 1024px resolution buckets.
A training assistant LoRA was produced by training a PEFT LoRA for directly onto the distilled Flux 3 Klein Image model for 30,000 steps on 30,000 real (15k illustrative and 15k photo) images over 8x H100 to break down the dual time distillation objective without introducing much bias into the resulting adapter.
Using that frozen LoRA, two photography LoRAs were trained against the full ~29k photo aesthetics dataset ( webshart/terminusresearch-photo-aesthetics) for 50,000 steps that demonstrate the ability of the training assistant to preserve the original dual time distillation objective.
Additionally, KohakuBlueleaf's KohakuFA (via https://github.com/SimpleTuner-io/KohakuFA-CuTe) was tested on two separate 50,000 step photography training runs to briefly evaluate the differences of the higher precision attention kernels.
I've put together a collection (https://huggingface.co/collections/SimpleTuner/flux-3-klein-image-lora-training) containing an experiment log book, a couple character LoRAs, and a 50,000 step diverse photo aesthetics finetune for the distilled variant in both 512px and 1024px resolution buckets.
A training assistant LoRA was produced by training a PEFT LoRA for directly onto the distilled Flux 3 Klein Image model for 30,000 steps on 30,000 real (15k illustrative and 15k photo) images over 8x H100 to break down the dual time distillation objective without introducing much bias into the resulting adapter.
Using that frozen LoRA, two photography LoRAs were trained against the full ~29k photo aesthetics dataset ( webshart/terminusresearch-photo-aesthetics) for 50,000 steps that demonstrate the ability of the training assistant to preserve the original dual time distillation objective.
Additionally, KohakuBlueleaf's KohakuFA (via https://github.com/SimpleTuner-io/KohakuFA-CuTe) was tested on two separate 50,000 step photography training runs to briefly evaluate the differences of the higher precision attention kernels.
bghira
updated a
dataset about 1 month ago
bghira
published a
dataset about 1 month ago
Post
109
Generated ~44 million Cosmos Reasoning captions for CC12M, now available here: webshart/cc12m-structured-captions
- webshart format, indexed with aspect bucketing metadata
- four captions per sample
- simple VLM output
- verbose caption
- a concise caption produced by passing the verbose caption through the model again to remove speculation and flowery language
- an ideogram4 style JSON schema
this used 8x H100 over about 6 days.
- webshart format, indexed with aspect bucketing metadata
- four captions per sample
- simple VLM output
- verbose caption
- a concise caption produced by passing the verbose caption through the model again to remove speculation and flowery language
- an ideogram4 style JSON schema
this used 8x H100 over about 6 days.
bghira
published a
dataset about 1 month ago
bghira
updated a
dataset about 1 month ago
bghira
updated a
dataset about 2 months ago
bghira
published a
dataset about 2 months ago
Post
2,746
I've been gradually recaptioning aged text-to-image datasets with better vision models! These datasets are also repackaged into the more modern webshart format (https://github.com/bghira/webshart) which includes built-in aspect bucketing and caption delivery.
The first two datasets are ready for use!
- webshart/terminusresearch-photo-anatomy
- webshart/terminusresearch-photo-aesthetics
"anatomy" is a bunch of human-centric images containing people holding or otherwise interacting with objects or positioned in complex ways.
"aesthetics" is a collection of visually striking images - high contrast, diverse colouration, and cinematic framing (among other factors).
These two datasets from 2023 were recaptioned with GLM 5.3 Flash via Ollama Cloud and Zhipu AI APIs, with a smaller portion run over 4x H100 with GLM 5.3 Flash in W4A16 precision - these outputs were checked by hand for quality, and the 4bit run was continued at a batch size of 64.
What datasets would you like to see recaptioned next? A better CC12M is on its way!
The first two datasets are ready for use!
- webshart/terminusresearch-photo-anatomy
- webshart/terminusresearch-photo-aesthetics
"anatomy" is a bunch of human-centric images containing people holding or otherwise interacting with objects or positioned in complex ways.
"aesthetics" is a collection of visually striking images - high contrast, diverse colouration, and cinematic framing (among other factors).
These two datasets from 2023 were recaptioned with GLM 5.3 Flash via Ollama Cloud and Zhipu AI APIs, with a smaller portion run over 4x H100 with GLM 5.3 Flash in W4A16 precision - these outputs were checked by hand for quality, and the 4bit run was continued at a batch size of 64.
What datasets would you like to see recaptioned next? A better CC12M is on its way!
- 3 replies
bghira
updated 3
datasets about 2 months ago
bghira
updated a
dataset 2 months ago
bghira
published 4
datasets 2 months ago
bghira
updated a
dataset 3 months ago