Repository navigation
Releases
Spectrum Krea 2 v0.2.2
Spectrum Krea 2 v0.2.2
The extrapolation distance guard can now be applied separately to the Scale-Locked planner and the high-resolution refinement passes.
Configuration
Keep extrapolation_guard_enabled=true and set extrapolation_guard_scope to one of the following values.
| Scope | Where the distance guard applies |
|---|---|
all (default) |
Planner and every high-resolution pass |
planner_only |
Planner, or the first segment of spectral-progressive sampling |
highres_only |
Every high-resolution pass, including later spectral-progressive segments |
Stage selection requires Scale-Locked Residual Diffusion v1.3.2 or later. Standalone samplers follow the master guard switch. Turning that switch off disables the distance limit everywhere. The other forecast checks still apply, and the session summary reports the stage and effective guard state.
Saved workflows retain their widget values. The new selector defaults to all, which preserves the previous guard behavior. Use ComfyUI Patcher to update to the released versions and remove the merged Spectrum #5 and Scale-Locked #21 / #22 overlays.
Validation
Native ComfyUI CPU tests verify guard routing and exact forecast counts on fixed schedules. One real-checkpoint ER-SDE run with planner_only used five guarded planner forecasts and one forecast on each of two unguarded refinement passes. This verifies the routing in that run. Broader image fidelity and reproducible GPU speedup remain unmeasured.
Spectrum Krea 2 v0.2.1
Spectrum Krea 2 v0.2.1 — corrected new-node defaults
This release aligns the node's initial controls with the tested Krea 2 Turbo non-bootstrap configuration. No sampling solver, forecast mathematics, numerical gate, or model interception logic has changed since v0.2.0.
Defaults for newly inserted nodes
preset=turbo,degree=1,ridge=0.01warmup=2,tail=1,window=1,growth=0bootstrap=falseto avoid the unvalidated one-point second-step feature holdextrapolation_guard_enabled=true,max_extrapolation_span_ratio=0.50sa_pece_forecast_policy=off;balanced_detailand other PECE skip modes remain opt-in until their own real-checkpoint quality testsstorage=vram,max_history=8,budget_mb=4096,debug=false
RAW users must select the RAW profile for RAW weights. The RAW example intentionally retains its independent 4/2 warmup/tail settings for more conservative initial comparisons. Existing saved workflows and imported examples keep their stored widget values; changing Python/UI defaults does not rewrite saved workflows.
Validation evidence and limits
A 2026-11-17 externally supplied paired Krea 2 Turbo ER-SDE sample at 3920 × 5888 pixels showed close visual composition. At 980 × 1472 downsampled resolution, the two JPEGs measured SSIM 0.9272 and PSNR 22.55 dB. Spectrum's planner used 9 real + 5 forecast evaluations in 14 steps; its four-step high-resolution pass performed zero forecasts due to the 0.50 extrapolation-distance guard. Total prompt durations (114.88 s Spectrum, 112.35 s native) were not directly comparable because model/text encoder/LoRA cache warmth differed. These data do not constitute a measured speedup or establish fidelity across prompts and seeds; the source JPEGs are not part of this release.
The default-only change is covered by schema/constructor/applied-wrapper agreement and saved-workflow explicit-override tests; bootstrap-specific tests explicitly opt in. CPU CI remains the structural release gate. See the validation record and GPU protocol. Real-checkpoint SA-PECE quality, CUDA precision, peak VRAM and representative end-to-end performance still require controlled evaluation.
Install or update through ComfyUI Patcher. This release contains the merged PR #2 interoperability code and retains the optional companion overlay requirements documented in the README.
Spectrum Krea 2 v0.2.0 (experimental)
Spectrum Krea 2 v0.2.0 — experimental sampler and patch interoperability
This prerelease extends Spectrum feature forecasting for native ComfyUI Krea 2 RAW and Turbo. It is intended for controlled checkpoint/CUDA image-quality and performance evaluation. Unsupported or changed execution contracts retain native sampling.
Changes since v0.1.0
- Added source-audited native forecasting paths for RES Multistep, Gradient Estimation, DEIS, iPNDM, iPNDM-V, stochastic/ODE ER-SDE, and SA-Solver (including predictor/corrector-aware PECE), alongside Euler and DPM++ 2M. Unreviewed solver options still fall back natively.
- Added qualified interoperability with Sol-Krea2 attention and Krea 2 Diff-Aid (text residual, values-only, keys-and-values). Patch-owner, config, attention-route and conditioning identities are rechecked during execution; changes cannot silently reuse incompatible feature history.
- Tracks Diff-Aid's effective gain per sampler call, including hard/partial/ramped sigma windows, and invalidates prior anchors when the gain changes. Session reports include bounded
diffaid_gainstraces. The companion Diff-Aidsigma_reference=model_sigma_maxoption can eliminate within-pass window transitions, but independently changes modulation and must be compared at matched seeds. - Separates native SA-Solver Adams/corrector state from ephemeral forecast endpoints. PECE policies
stable_start,balanced,max_speed, andbalanced_detailare experimental; the defaultoffremains the baseline. - Adds invocation-scoped
FORCE_NATIVE_ALL_SAMPLERSfor planner-only quality isolation, including ER-SDE and Euler. The pre-existingFORCE_NATIVEcontinues to apply only to SA-Solver. Both scopes restore their previous value on exit. - Adds short partial-denoise regression coverage, split-CFG attention-route isolation, native-only parity, callback/exception cleanup, forecast guard boundaries, detailed fallback reasons and build validation.
Installation and companions
Install or update this repository through ComfyUI Patcher and restart ComfyUI. The Spectrum PR #2 changes are included in this release; do not apply PR #2 on top of v0.2.0.
Sol-Krea2 support was introduced by Sol-H3 PR #44, which is merged. Krea 2 Diff-Aid requires the Diff-Aid PR #12 overlay until that PR is merged. Planner-only native controls in the companion Scale-Locked sampler require Scale-Locked PR #22, stacked on PR #21. These external patches are optional; the base Krea 2 workflow does not require them.
The default Sol dense_evaluations=1 can consume the only eligible forecast in a four-step high-resolution stage. Changing it to zero changes the attention operator from the first evaluation, and may change image quality. Short stages also restart Spectrum history. No incompatible Diff-Aid gain or attention route may share forecast anchors.
Verification
- GitHub Actions at PR #2 head
80ec4fba98f6def731f37d8930e0281ed7f2cc98: 379 CPU tests passed in each Python 3.12 current/legacy sampler lane; Python 3.10/3.13 lanes passed. - 40 native source contracts matched. Full ComfyUI CPU integration passed for both sampler-container variants; PR validation records 82 integration cases per variant.
- Packaging and import checks passed. The 14 inline CodeRabbit review threads were resolved.
- The release metadata and documentation update contains no numerical sampler changes; release-commit CI runs independently after the PR head is updated.
Limits and acceptance
No real Krea 2 checkpoint CUDA quality or performance validation is established. CPU scheduling, bitwise native/all-actual parity on small weights and counts of forecasts do not establish decoded image fidelity, attention-kernel correctness, peak VRAM or acceleration.
Use matched seeds to isolate changes. For Diff-Aid partial windows compare (a) schedule_start with Spectrum off, (b) model_sigma_max with Spectrum off, and (c) model_sigma_max with Spectrum on. Keep sa_pece_forecast_policy=off for baseline comparisons. Follow the GPU validation protocol before recommending accelerated defaults for production.
Spectrum Krea 2 v0.1.0 (experimental)
Spectrum Krea 2 v0.1.0 — GPU-validation prerelease
Initial native ComfyUI Krea 2 integration of Spectrum feature forecasting for RAW and Turbo models. This is a prerelease intended for paired image-quality and performance validation.
Features
- Spectrum Apply Krea 2 as a MODEL → MODEL node after the final model/LoRA loader.
- Capture the native generated-image transformer features before the output layer, fit causal Chebyshev ridge history and project forecasts through the native current-timestep output head. No additional transformer or model weights.
- RAW and Turbo profiles; Euler and DPM++ 2M source-contract gating, bounded forecasting, optional one-point second-step bootstrap, per-invocation session ownership, stream isolation and native fallback.
- CPU-tested reference conditioning, CFG and batch packing, odd spatial shapes, image-to-image denoise, native LoRA weights, singleton-frame 5D latents and callback/exception cleanup.
- Matched UI and API example workflows, GPU benchmark runner and paired-image comparison utilities.
Verification
- GitHub Actions: 136 CPU tests passed in the latest Python 3.12/current-container lane. Python 3.10/3.13 and an earlier sampler-container lane also passed.
- 16 full native ComfyUI CPU integration cases passed for each of the audited sampler containers using small random test weights.
- CPU tests establish routing, native-output parity for disabled/all-actual execution, lifecycle safety and forecast execution, not production image fidelity.
Limitations and validation requirement
Real-checkpoint image quality, CUDA precision and speed, peak VRAM and actual acceleration remain unverified. No published default is claimed to be a measured quality/speed optimum. Forecast counts alone are not speedup data.
Start with the conservative paired RAW (52 steps, CFG 4.5) or Turbo (8 steps, CFG 1) Euler/simple examples at 1024×1024. Compare the same seed, model, text encoder, VAE and LoRAs with and without Spectrum. Inspect composition, text, anatomy and small detail and measure wall time and memory. See GPU validation.
Use ComfyUI Patcher to install the release and restart ComfyUI. Unsupported samplers, modified transformer contracts and unsupported conditioning paths fall back to native execution.