feat: AnimateDiff img2video via --init-img

[fszontagh]

Summary

Adds img2video mode to AnimateDiff (SD 1.5): a pre-rendered init image is VAE-encoded, replicated across N frames on the batch axis, and used as the starting latent for the diffusion loop. Frame 0 preserves the init image's quality; subsequent frames drift slightly via the motion module.

Follows PR #1784 (AnimateDiff SD 1.5 v2/v3 motion modules).

Related Issue / Discussion

None.

Additional Information

Sample

Left: init image rendered separately with -M img_gen. Right: 8-frame -M vid_gen -i init.png --strength 0.75 output. Character identity, dress, background, sharpness all preserved. Motion module adds subtle drift (hair, lighting, breathing).

Verification command

sd-cli -M img_gen \
    --model realisticVisionV60B1.safetensors \
    -p "..." -H 512 -W 512 -s 42 --steps 25 \
    -o init.png

sd-cli -M vid_gen \
    --model realisticVisionV60B1.safetensors \
    --motion-module mm_sd15_v3.safetensors \
    -i init.png --strength 0.75 \
    --cfg-scale 7.0 --sampling-method euler --scheduler karras \
    -H 512 -W 512 --video-frames 8 --steps 25 -s 42 \
    -p "..." -o out.avi

How it works

Two changes in src/stable-diffusion.cpp:

1. prepare_image_generation_latents: when animatediff_num_frames > 1 and init_image_tensor is non-empty, replicate the single-frame encoded init latent across the N-frame batch axis via sd::ops::slice_assign. When no init image is provided, the existing T2V zero-fill path runs unchanged.
2. generate_animatediff_video: plumb sd_vid_gen_params->init_image into img_gen_params.init_image so -M vid_gen -i <path> reaches the latent preparation path.

13 insertions + 2 deletions in src/stable-diffusion.cpp. No public API changes.

Checklist

• I have read and confirmed this PR follows the contribution guidelines.