SnapRef

SnapRef v2a stable. Character reference, multiple views of one character, per-reference weights, and Character mode for prompt-driven styling.Appearance-tag recall improves slightly, while our qualitative assessment shows substantially better reference following than v1a. A remaining issue is that the model can struggle to reproduce facial details when the face in the reference image is very bright.Download <code>snapref_v2a.safetensors</code> and the attached <code>snapref_v2a_comfyui_node_and_workflows.zip</code>. The ZIP includes the SnapRef custom node and six workflows. Start with <code>workflows/snapref_v2a_simple.json</code>. SnapRef requires its own node; a regular LoRA loader will not work.Base: Anima v1.0. Weights: CircleStone Labs Non-Commercial License v1.2; license and attribution are included in the ZIP.Optional FP8 download: <code>snapref_v2a_fp8.safetensors</code> is approximately 195 MB (about 49% smaller). Select it in the same SnapRef node. It uses E4M3FN matrices with per-row scales; inference still computes in BF16. Generation results can differ slightly from the original weights.

Files with SHA256: 1f54971e14bc1f56569acd99304d774af41857273fc4aee167cf2d50a7248e92