UPSTREAM PR #1296: feat: add Anima support

[loci-dev]

Note

Source pull request: leejet/stable-diffusion.cpp#1296

For leejet/stable-diffusion.cpp#1245

./build/bin/sd-cli --diffusion-model models/anima-preview.safetensors --llm models/qwen_3_06b_base.safetensors --vae models/qwen_image_vae.safetensors -p "a cute cat" --fa -v -H 1024 -W 1024 --cfg-scale 4

Download model: https://huggingface.co/circlestone-labs/Anima/tree/main/split_files

Output eg:

[loci-review]

Overview

Analysis of 49,755 functions across two binaries reveals mixed performance impacts from adding ANIMA model support. Modified: 122 functions, new: 1,454, removed: 4, unchanged: 48,175.

Binaries analyzed:

• build.bin.sd-server: +1.458% power consumption (518,798 → 526,362 nJ)
• build.bin.sd-cli: +1.574% power consumption (483,665 → 491,278 nJ)

The single commit "add anima" introduces a new diffusion model architecture through additive changes (774-line anima.hpp, model detection logic). Performance changes stem primarily from compiler optimization artifacts rather than algorithmic modifications.

Function Analysis

STL Vector Accessors - Improvements:

• std::vector<gguf_kv>::end() (sd-server): Response time 263.83ns → 80.54ns (-69.47%), throughput 243.07ns → 59.78ns (-75.41%)
• std::vector<std::thread>::end() (sd-cli): Response time 265.16ns → 81.87ns (-69.12%), throughput 243.07ns → 59.78ns (-75.41%)
• std::vector<ggml_backend_reg_entry>::cbegin() (sd-server): Response time 264.07ns → 83.25ns (-68.47%), throughput 243.31ns → 62.49ns (-74.32%)

These improvements benefit model loading and backend initialization through better compiler code generation for simple template types.

STL Vector Accessors - Regressions:

• std::vector<std::pair<...>>::end() (sd-server): Response time 80.53ns → 263.84ns (+227.65%), throughput 59.77ns → 243.08ns (+306.70%)

Complex nested template types suffer from optimization challenges, affecting AnimaConditioner initialization.

Model Detection Logic:

• sd_version_is_dit() (both binaries): Response time +9.07% (~31ns), throughput +10.06% (~9ns)

Functionally justified regression from adding ANIMA classification check, enabling DiT optimizations for the new model.

Memory Management:

• std::shared_ptr<T5CLIPEmbedder>::_M_destroy() (sd-cli): Throughput 293.74ns → 105.03ns (-64.24%), response 495.60ns → 307.53ns (-37.95%)
• std::make_shared<WanModel> (sd-server): Throughput +43.38% (+41ns), but response time +0.05% (468ns out of 938µs)

Construction shows throughput regressions confined to function prologues; destruction improves significantly.

Other analyzed functions (std::swap, thread::joinable, file I/O operations) showed minor regressions (8-59% throughput increases, 9-76ns absolute) attributable to compiler optimization variations rather than source changes.

Additional Findings

Performance changes are isolated to initialization and utility code; inference hot paths remain unaffected. The analyzed functions do not involve GPU kernel execution—impacts are limited to model loading, backend registration, and object lifecycle management. The 1.5% power consumption increase is negligible compared to actual inference costs. ANIMA integration successfully leverages existing DiT optimization infrastructure (EasyCache, CacheDIT) with minimal overhead.

🔎 Full breakdown: Loci Inspector
💬 Questions? Tag @loci-dev