Module stable diffusion Future

Hinweis: Vage Einträge ohne messbares Ziel, Interface-Spezifikation oder Teststrategie mit `` markieren.

Future Enhancements — Stable Diffusion Plugin

Scope

Planned enhancements beyond v2.0.0. Core implementation in src/stable_diffusion/sd_plugin.cpp, src/stable_diffusion/sd_prompt_sanitizer.cpp, src/stable_diffusion/sd_config.cpp.

---

Design Constraints

• IImageGenerationBackend interface must remain stable; new capabilities added as optional methods with default implementations.
• Content-policy screening (SDPromptSanitizer::isAllowed) must be called before every ISDGenerator::generate call — no code path may bypass it.
• Provenance stamps (plugin_version, generation_timestamp, prompt_hash) must always be applied by SDPlugin, never delegated to the generator.
• Stub mode (no model file) must continue to produce a valid GeneratedImage with success=true after any enhancement.
• FNV-1a hash for prompt_hash must remain the same algorithm across versions to keep hash values stable in audit logs.

---

Required Interfaces

Interface	Consumer	Notes
ISDGenerator::applyLoRA(lora_path, scale)	LoRA hot-loading	Implemented; request-time validation required
SDGenerationConfig::control_* fields	ControlNet path	Implemented; control-image shape/strength validation required
GeneratedImage::perceptual_hash	Dedup/Audit	Implemented; non-fatal if unavailable

---

Delivered Since Initial Baseline

• ControlNet request fields and validation are implemented in SDGenerationConfig/Img2ImgConfig and enforced in SDPlugin.
• LoRA hot-loading request path (applyLoRA) is implemented and validated.
• Optional GeneratedImage::perceptual_hash metadata is implemented with non-fatal fallback.
• Model integrity check (model_sha256) is implemented during initialize().
• Dimension guards are implemented for request and output validation.

---

Planned Features

1. Performance/Hardening Follow-ups (Target: Q3–Q4 2026)

Problem: Roadmap Phase 5 contains open hardening items.

Solution: Keep benchmark and concurrency-audit deliverables:

• benchmark target added: bench_stable_diffusion_release_gates for time-to-PNG (stub + in-memory proxy); publish real-model baseline next
• thread-safety audit target added in SDPluginRealBackendE2ETests; publish CI/runtime evidence next

Tests: Benchmark harness entry + stress tests gated for SD backend availability.

---

2. Real-backend E2E Automation (Target: Q3 2026)

Problem: Focused unit tests validate logic, but no dedicated stable_diffusion E2E gate exists for real model execution in CI.

Solution: Maintain opt-in E2E target that runs with THEMIS_ENABLE_STABLE_DIFFUSION=ON, a test model path, and SHA-256 verification enabled.

Tests: one text2img and one img2img E2E case asserting valid PNG output and metadata.

---

Security / Reliability

• Any new image decoder path must validate dimensions before allocation to prevent integer-overflow-based heap attacks.
• Model integrity verification (SHA-256 policy) is enforced when model_sha256 is provided.
• Any future batch override in SDCppGenerator must preserve per-prompt SDPromptSanitizer::isAllowed() checks.

---

ThemisDB 1.9.0-beta · Home · Module-Index · GitHub · Issues