RBG Smart Seed Variance — major update 2.2.0: Krea2 support, variance scheduling, interactive token protection, and direction shift library expansion and more..
- Add Krea2 (SingleStream) model type with band-aware noise handling across
the 12-layer MMDiT tap stack, including a dedicated base-embedding
rebalance pass to boost expression, attention detail and prompt following.
- Add ERNIE-Image to model-specific adjustment presets, alongside
refined multipliers for Flux, Chroma HD, SDXL and Wan2.2.
- Add interactive token protection inspector (JS): click-to-toggle individual
tokens directly on a live visualisation of the protection mask, replacing
manual range typing as the primary workflow.
- Add "Custom Regions" and "Random Regions" protect modes — custom accepts
explicit token/range syntax (e.g. "0-5,15-20"), random seeds a procedural
protection mask.
- Add live protection mask feedback from backend to frontend via onExecuted,
so the inspector reflects the actual tokens used in the last run rather
than a static preview.
- Add Variance Schedule (Composition Lock 🔒) — control how noise intensity
changes across the sampling timeline with 5 modes: constant, decreasing
(fade out), step_cutoff (hard switch at a step), hard_lock (zero variance
until cutoff), and tiered_release (multi-phase staged unlock)
- Add cutoff_step, total_steps, and cutoff_strength controls — map a step
number to a percentage of the sampling timeline and set the noise
multiplier applied after the cutoff point, so composition can be locked
early while detail variance continues later.
- Expand direction shift library from the original artistic set to 26+
patterns, including Face-Variance Expansion, Visceral Expression & Grit
(Krea2-tuned), Semantic Drift, Structural Lock, Cinematic Framing, Texture
Lift, Landscape Depth, Group Diversity, Expression Variance, Motion Blur,
Microscopic, Cosmic, and Bone Anatomical Coherence.
- Add dynamic widget visibility — protect_regions field now shows/hides
automatically based on the selected protect_mode.
- Fix word-index vs BPE-index mismatch in token protection by adding
bidirectional BPE-to-word mapping, so protected ranges align correctly
with actual tokenizer output instead of word-level approximations.
Ramon Guthrie