Handle multi-control control images when encoding prompts

[ojasaar]

Summary

Fixes #441

• ensure multi-control models load every available control image when encoding prompts
• handle batches where control tensors are stored as per-sample lists and map them into the shapes Qwen expects
• keep behaviour unchanged when no control images are present

Testing

• python3 -m compileall extensions_built_in/sd_trainer/SDTrainer.py toolkit/models/base_model.py
• sampling + training with qwen image 2509 on a dataset with 3 control images

[lucusmax]

im still getting pixel value errors after merging - training without samples seems to be working. is there any update coming for this? thanks

[ojasaar]

im still getting pixel value errors after merging - training without samples seems to be working. is there any update coming for this? thanks

I've tested it with having all 3 control datasets. How many are you using? Do you get errors during sampling or training?

[lucusmax]

I got it to work - through remembering to add control images to the sample prompts....forgot about that. Adding one control image to each of the sample prompts allowed it to go through - running now and seems to be training well.
thanks for all your help.

[Tristan-mc-q]

hey @ojasaar im still getting errors using 2 different control images on a 143 image triplet dataset. All images are labeled the same and I even disabled sampling to see if I could get things running. Im running on an h100 on runpod. When inspecting further I found Preprocessing shows 100% 143/143 and "Found 143 images", but only ~23 UUID files get bucketed in /data/images

the "preprocessing" phase doesnt seem to be finishing completely. It's only processing ~23 out of my 143 images, creating temporary files with random names (like "03868a7c-...png") that don't match my original numbered files (001.png, etc.).

Edit: Hold on didnt properly test your merge. Let me double check this is solved

[Astroburner]

Summary

Fixes #441

• ensure multi-control models load every available control image when encoding prompts
• handle batches where control tensors are stored as per-sample lists and map them into the shapes Qwen expects
• keep behaviour unchanged when no control images are present

Testing

• python3 -m compileall extensions_built_in/sd_trainer/SDTrainer.py toolkit/models/base_model.py
• sampling + training with qwen image 2509 on a dataset with 3 control images

Well, I have 1 dataset and just one set of control images.
you said you was using 3 sets of control images?

It dident fix my issue.

My plan was a simple switch LoRA for qwen2509.
Also i uploaded for sampling just one image because i dont need 3.

Do you have other ideas how to solve it?