DalleWebms

birds, mega model

maid knight, mega model

rooms, mega model

More: https://github.com/TabuaTambalam/DalleWebms/blob/main/rooms.md

interpo "superman", mini model

Outpainting:

without user select

• vq.param for NCNN here: https://github.com/TabuaTambalam/vqqncnn can decode width-free image like those above. vq_vert.param can do height-free decode.

More samples here: https://github.com/TabuaTambalam/DalleWebms/tree/main/docs/still

• original seqs of all those samples: https://github.com/TabuaTambalam/DalleWebms/releases/download/0.1/seqs_outpainting.zip

Interactive Outpainting:

• Notebook here: https://colab.research.google.com/github/TabuaTambalam/DalleWebms/blob/main/min_dalle_interactive_hacky.ipynb
• Results from rudalle's ESRGAN & guided diffusion (now inside the notebook above) here: https://github.com/TabuaTambalam/DalleWebms/blob/main/rudallestuff.md

• Early results when attention_state not duplicated yet: https://github.com/TabuaTambalam/DalleWebms/blob/main/notfixAS.md

with crop (overall color tone will alter):

mini-model:

Infinite similar gen:

• With min_dalle_interactive_hacky.ipynb, run 'Make a prompt' first with your prompt,
• move to 'Infinite similar gen', run the cell the first time.
• choice an initial image with candidate_select, run it the second time,
• now the infinite gen thread started, click the showp(-1) cell when you see the ozv.bin file size growth.
• to stop the infinite gen thread, rename once.txt to -.txt, when you see once.txt re-appears, the thread is stopped.
• results are stored in ozv.bin as original seqs.
• examples:

• A good initial image (well, good for the model, not human) will bring good results afterward. Those below kept top two rows locked in token space:
• (seqs: https://github.com/TabuaTambalam/DalleWebms/releases/download/0.1/ozv_lock3_fefaces.7z , 8880 images as it decoded)

• more: https://github.com/TabuaTambalam/DalleWebms/blob/main/docs/still/coomerbait/locktop2a.md
• https://github.com/TabuaTambalam/DalleWebms/blob/main/docs/still/coomerbait/locktop2b.md
• https://github.com/TabuaTambalam/DalleWebms/blob/main/docs/still/coomerbait/locktop2c.md
• Also locktop2d.md, locktop2e.md ... try until you get 404'd

Height-free decode:

About Min_dall_singlfileGPU.ipynb

• Most codes came from https://github.com/kuprel/min-dalle , merged into single notebook, with following modifications:
• Use meta device for nnModule initializing, read more here: https://github.com/FrancescoSaverioZuppichini/Loading-huge-PyTorch-models-with-linear-memory-consumption
• Hardcoded to fp16 (the checkpoint itself is fp16, no need to upconv them to fp32 unless for speed reason)
• Partially-locked generation.
• Use ncnn VQGAN for decoding & interpolation, those animations above produced by ncnn VQGAN.
• Original seqs of those animation above is https://github.com/TabuaTambalam/DalleWebms/releases/download/0.1/ozv.bin
• colab link https://colab.research.google.com/github/TabuaTambalam/DalleWebms/blob/main/Min_dall_singlfileGPU.ipynb