Attn2 SVD Prompt Injection

"Attn2 SVD Prompt Injection" is the odd one out in this pack: it's aimed at Stable Video Diffusion, not SDXL. Same core idea - override what specific blocks of the UNet get to look at - but pointed at the video model's block layout.

SVD's UNet is structured like a diffusion UNet with video-specific conditioning paths, and this node exposes those as conditioning ports: time_embed and label_emb (the micro-conditioning stages), plus input_blocks_0 through input_blocks_8, middle_block_0 through middle_block_2, and output_blocks_0 through output_blocks_8. There's an all port that acts as the default for every block, a weight, and the usual start_at/end_at denoise window. Feed a conditioning into any of them and the patch swaps it in while that block runs.

One honest note: in SVD's case that conditioning isn't a text prompt. SVD drives its cross-attention with CLIP image embeddings from a conditioning frame, not with words, so "prompt injection" here really means "conditioning injection." Your conditioning input comes from the image-encoding side of the video pipeline, not a plain text prompt. The display name inherited the pack's brand; the mechanism is what it always was.

Should you use it?

Only if you're actively tinkering with SVD video generation and want the same block-level surgery you'd do on SDXL. If you're not generating video, this one is not for you - it does nothing on a still-image model.

And there's a real caveat before you build a workflow around it: the pack's current code concentrates on the SDXL and SD1.5 nodes. This SVD variant appears to have lived in an earlier revision of the repo. If you install today's pack and SVDPromptInjection isn't in your node list, that's likely why - grab an older revision or check whether the repo has caught up before you count on it. It's the least-maintained node in the pack, and it shows.

Installing

Identical to the rest of the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/DataCTE/prompt_injection

or ComfyUI Manager → search "Prompt Injection", then restart. No pip dependencies, no model downloads.

Gotchas

• No conditioning connected (and no all) = silent no-op, same as every node in this pack.
• If the node isn't in your node list after a fresh install, see above - the current repo's lineup is SDXL/SD1.5.
• The block names are SVD-specific (time_embed, label_emb, input_blocks_N, …). Feed it into an SDXL sampler expecting text-attention blocks and it won't do anything.