daily vlog of tracer (overwatch)

https://redd.it/1t706ss
@rStableDiffusion

I finetuned Qwen3-1.7B to imitate original Z-Image text encoder. 21% less VRAM

https://redd.it/1t71hvm
@rStableDiffusion

LTX 2.3 ID-LoRA with First-Last Frame

The official ComfyUI ID-LoRA workflow for LTX-Video 2.3 only supports first-frame conditioning out of the box, which limits how much control you have over character motion and pose. I wanted to add last-frame support with minimal changes to the original — no restructuring, no new samplers, just surgical node edits. You can grab the modified workflow here.

What was changed:

The default workflow uses LTXVImgToVideoInplace (comfy-core) for image conditioning in both the low-res and high-res sampling passes. This node only handles a single frame at a fixed position. The fix was to swap both instances out for LTXVImgToVideoInplaceKJ from KJNodes, which supports multiple images at arbitrary frame positions in a single call.

Concretely:

1. Added last-frame preprocessing — two new nodes mirror the existing first-frame preprocessing pipeline: a ResizeImagesByLongerEdge (1536px) followed by LTXVPreprocess. These feed the last-frame image into both sampling passes.
2. Low-res pass — The LTXVImgToVideoInplace node was replaced with LTXVImgToVideoInplaceKJ configured for 2 images: first frame at position 0, last frame at position -1, both at strength 0.7. One node, both frames conditioned simultaneously.
3. High-res pass — Same conversion applied to the conditioning node after LTXVLatentUpsampler. Both frames re-conditioned at strength 1.0 so the last frame gets sharpened in the upscale pass just like the first frame. Without this step the last frame came out noticeably blurrier.
4. New subgraph input — A last_frame image input was added to the workflow's subgraph, wired to a LoadImage node on the canvas.

That's it — 2 node type swaps, 2 preprocessing nodes, 1 new input. Everything else (sampler, audio conditioning, LoRA stacking, the upscale pipeline) is untouched from the official Comfy Cloud release. Let me know if you have any questions. Cheers!

https://redd.it/1t71x0r
@rStableDiffusion

Why is it that 3 years old SDXL is still the best base for porn checkpoints, where the best ones on civitai produce materially better images than the z image or flux porn checkpoints in terms of realism and skin texture?



https://redd.it/1t71cs5
@rStableDiffusion

Continuous-Time Distribution Matching: A new SOTA method for step distillation.

https://redd.it/1t76p1t
@rStableDiffusion