yada

[yada @yada_cc]

@imgggalvin \强强/

[yada @yada_cc]

@teortaxesTex It's just a heavier Typora for me, like sublime vs. vscode

[yada @yada_cc]

Just finished the #HuggingFace Agents course, I think it's a good overview if anyone’s looking into agentic workflows🤗

[yada @yada_cc]

@Sarahhuniverse irrational numbers mod 1 be like

[yada @yada_cc]

I'm un-AIbroing

[yada @yada_cc]

I'm about to pull off something very funny

[yada @yada_cc]

I'm about to pull off something very funny

[yada @yada_cc]

@_catwu heard about claude code on max plan has less context length than pay-as-you-go, is that true?
https://t.co/YuzLEw5DvY

[yada @yada_cc]

@signulll Given infinite time one can surely learn into anything. But in reality you'd have to make tradeoffs. Free time is the real flex.

[yada @yada_cc]

@djcows need better instructions

[yada @yada_cc]

If you think about it... sufficient statistics are basically embedding models with zero reconstruction loss.
It's been compression all along, for 103 years.

[yada @yada_cc]

Fresh Danbooru 2025 metadata, scraped up to Apr 16
https://t.co/wkvXrBt0HU

[yada @yada_cc]

Progress / experiences so far on #hidream lokr trainings, before I forget:

vram (bf16, 1024px):
- 24G for square images single aspect ratio bs1 on dev (not sure)
- 32/48G for full, around bs1
- 80G for aspect buckets + bs8

training:
- use SimpleTuner
- use EMA (doesnt take vram, and it really helps)
- lion 3e-5 if bs1; larger than 1e-4 for bs64
- enabling arb if you can (++ vram, also needs more pics)
- train on full (HiDream-ai/HiDream-I1-Full); dev won't work
- add repeat if not enough images to fill buckets
- remember to modify `lycoris_config.json`
- cosine scheduler seemed broke for me, use polynomial for now
- add `--override_dataset_config` for hyperparam tests

model:
- have descent understanding of tags and captions, can write text too
- have a base 'anime style' with good anatomy (that is slightly better than flux)
- if trained on i1-full, use full for inference too (dont use dev)
- supports 128 tokens MAX on all 4 encoders,
- ^ extending llama tokenizer to 256 taken take 10G extra (63 -> 73G) for inference loosely works, but have color issues

captioning:
- add a trigger word prefix (or eval breaks)
- kinda works with danbooru tag captions

speed:
- 14s per image 30 steps on 1xH100
- 5h30min for 1200steps bs8*acc2, lokr factor16 1024-arb-bf16 on 4*H100

I will update under replies if got anything new

[yada @yada_cc]

aaaand.... #OpenAI Monday's gone.

It was a good run though, saved the system prompt here in case anyone wants to recreate https://t.co/qenfdOfovw

[yada @yada_cc]

HunyuanVideo PixelArt Lora

(should had released it on Janurary but still)
-> https://t.co/dVmiWkna59

[yada @yada_cc]

I should start posting more

[yada @yada_cc]

@ctjlewis @ollama Which confuses a lot of people, because anything other than r1:671b are really just sft distills instead of the real thing, yet they're all labelled as 'r1' on ollama.

[yada @yada_cc]

r1 is based