YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
NEURONAUT diffusion uSAE — retained research interims
Work in progress, not a validated release. These two private repositories preserve actual retained checkpoints and evidence, including incomplete and failed stages.
inventory.json resolves every retained logical file to repository, path and SHA256. Identical activation caches and replayed checkpoints are stored once and referenced; preseeded caches are not evidence that a run executed. Overwritten historical checkpoint versions and unretained remote artifacts are not claimed recovered.
Bulk artifacts are retained here on Hugging Face. Verified local research copies and project model caches were evicted on 26 October at the user's request; local absence is deliberate. New backup transport passes artifacts through bounded host RAM, verifies the remote content, and keeps compact local receipts. It does not maintain a laptop bulk mirror. Cleanup reports and the remote-path catalogue are under retention/.
Current state — 26 October 2026
neuronaut-unified-1025 completed: seven native generator configurations, 8,192 shared hierarchical updates, 140 held-out native factor tests, measured train-only dictionaries and portable inference exports. All 3,182 terminal unique artifacts were independently remotely content/size verified. All 140 native no-op controls are exact. The run used 3.894 useful L4 hours. Shared coordinates show spatial co-firing above within-image controls, but this does not establish a semantically validated universal dictionary.
The matched no-reference hierarchy control and flat10,240-coordinate BatchTopK MSE comparison in neuronaut-controls-1026c both completed, including native tests, empty-prompt palette trials, measured indexes, portable exports and full147,170-term candidate-label retrieval. All2,176 terminal unique artifacts were independently remotely hash/size verified. The run took53.71 minutes and released its L4; fresh server status confirms no active sessions. The no-reference fit took259.91seconds, the flat fit208.74seconds.
Removing an inconsistent frozen-probe penalty improves SDXL self R² from0.136 to0.337 and native→DINO from−0.238 to0.109 on the same1,024 validation rows. FlatBatchTopK self R² is0.423, while the hierarchy has better DINO→native prediction. Native boat/background edits and shared zebra-striped exemplars are concrete positive findings. Empty-prompt palette transfer remains weak, and retrieved labels include lexical errors (a water-surface feature called “watery-eyed”). These findings do not establish a validated universal painting dictionary. The flat comparison changes capacity, loss and structural constraints together; it cannot isolate hierarchy alone.
The private CPU duplicate and authorized GPU Space now have a saved-experiment Diffusion Lab tester with separate hierarchy/control namespaces, candidate-name lookup and palette trials, placed last after Method. Hosted comparison, measured exemplars, compact export and existing state-sync callbacks passed without a ZeroGPU allocation. This is not yet a live multi-model editor. The original CPU Space remains untouched. See CURRENT.md and evidence/coco_cohort/hosted_tester_checks_20261026.json.
UNIFIED_METHOD_20261025.md defines the source-grounded extension. Historical checkpoints, unsuccessful controls and sources remain preserved; the following dated sections describe earlier stages rather than superseding this status.
Current diffusion alignment work —25 October2026
Interpretability and cross-model concept alignment are primary. Editing and morphing are downstream demonstrations; NMF is a feature-grouping and mapping instrument throughout the work.
Completed:
- Full eight-scene causal surveys at2354 sites across SDXS, DreamShaper, SD-Turbo and SDXL-Turbo;321 produced artifacts remotely verified.
- Twenty-scene surveys of all whole attention/ResNet updates:18 SDXS,15 DreamShaper,38 SD-Turbo sites. Native heatmaps, exact no-op controls, LPIPS and DINO readings retained.
- Six SDXS block-specific Surkov-style TopK10 dictionaries trained on measured updates from3072 COCO-generated scenes,32 native patches/image. Separate256-image validation and test partitions. All six final weight hashes independently verified on this private model repo.
| SDXS site | Validation R² at1024updates |
|---|---|
| down_blocks.0.attentions.0 | 0.5475 |
| down_blocks.2.attentions.0 | 0.5191 |
| up_blocks.0.attentions.0 | 0.5313 |
| up_blocks.0.attentions.1 | 0.5121 |
| up_blocks.2.attentions.0 | 0.4608 |
| up_blocks.2.attentions.1 | 0.8511 |
Native suppression/enhancement grids cover20 fixed scenes per site and transfers to related generators. The deep down-attention lesions change object identity, number and composition; final-stage lesions can destroy image formation. Reconstruction score alone does not choose the semantic junction.
The subsequent native capture and joint fit did execute:1280 images per branch, then6144 updates logged in358.2seconds. Step3072 is the last fully restorable checkpoint. All three6144 branch weights and optimizer payloads are retained, but final RNG/manifest and six native-validation artifacts did not mirror before guard release. Step6144 weights are partial research interims, not a complete resumable checkpoint. The underlying worker termination cause is unresolved. Step3072 cross-model reconstruction was poor despite positive self-reconstruction. DINO patch-pair inspection shows different object presence, scale or viewpoint, but shared backgrounds/textures can remain valid features. Foreground mismatch is not an automatic rejection rule. Assess component-level shared information before changing the matching policy; a reciprocal matcher is prepared as a diagnostic only.
Completed inference-only block study
Run neuronaut-blocks-1025 covers167 attention/ResNet units across seven configurations,20 fixed scenes each, native residual gains0/1/2 and native512px output-difference maps. All167 gain-one controls exactly reproduce baseline. No SAE fitting participates. SDXS/DreamShaper/Anime/Sketch/SD-Turbo native20-image conditions take roughly4.5–4.9seconds each; SDXL-Turbo5.3seconds, native25-step SD2.1 31.7seconds. Both suppression and amplification are retained per block.
The deep SDXS down-attention changes grouping and subject relationships; first up-stage attention changes identity/appearance; late ResNet affects sharpness. The largest catastrophic image-formation lesion is distinct from the composition junction. Native SDXL down2.1 and the two adjacent SD2.1/SD-Turbo deep attention units give a functional reference; the joint teacher-unit probe is prepared but has not executed.
Independent SDXS feature4215 at up_blocks.0.attentions.1 was investigated through eight native replays and gated-code0x/4x interventions. Maps/effects support lower-face/neck/chest fur, with strong dog responses and weaker bear/cat responses. This is a retained TopK10 coordinate, not a shared BatchTopK ID. See MECHANISTIC_FINDINGS.md, the portable runs/neuronaut-blocks-1025/results/block_response_viewer.html, native feature panels and evidence/coco_cohort/feature4215_dog_bear_cat.webp.
Feature-to-concept membership remains many-to-many. Following Goodfire's NMF formulation, a global SAE feature may contribute to several per-image factors; colors, textures and backgrounds are valid shared content. Different foregrounds do not alone invalidate patch correspondence. See the overlapping-concept discussion and source links in MECHANISTIC_FINDINGS.md.
CAME and AdamW kernels were measured rather than assumed faster: fused AdamW317updates/s, default AdamW126, CAME67. CAME halves persistent optimizer state but has a higher temporary peak. These are kernel measurements, not training convergence or full-loop speed. Fresh CUDA fits now select fused AdamW; CAME is explicit. The next shared fit follows actual correspondence and signal-definition checks, with native exemplars and causal effects used to interpret Matryoshka/NMF groupings. Spaces remain unchanged during representation research.
COCO captions come from the official public2017 annotations.13,824 image-disjoint inputs and all five native tokenizer views were checked. Captions drive inference; SAE targets are neuronal activations. Exact source/weight/corpus hashes and run states are retained.
Earlier fixed-per-patchTopK research retained
31 eligible Transformer2D additive-update sites were screened across SDXS, SD-Turbo and Dreamshaper. The composition candidate for the primary pair is down_blocks.2.attentions.0, native 16×16, width 1280. This is a scoped screen, not a claim of global optimality.
The dictionary has 10,240 coordinates. Sparse codes are trained on frozen generator activation tensors; text and noise are stimuli only, never SAE labels or targets. The study compared K=32/128/256 and aligned versus independently trained, train-only feature-matched controls. Five complete capacity-study artifact sets survive; the independent K256 final set was not retained.
Selected K128 branches were replayed with bit-identical weights. On 768 fresh confirmation stimuli:
| Pair | SDXS own R² | SD-Turbo own R² | SDXS → Turbo R² | Turbo → SDXS R² |
|---|---|---|---|---|
| Aligned K128 | 0.7979 | 0.7411 | 0.6140 | 0.6319 |
| Independent K128, train-only matching | 0.8249 | 0.7658 | 0.1537 | 0.1558 |
These results establish reconstruction and cross-prediction, not semantic universality. Native image-preservation controls and their raw reports are retained.
Incomplete stages and failures
- K10 and K32 utilization failed the declared gate; retained as negative results.
- One exact replay was stopped by an overly strict comparison of reordered Diffusers bookkeeping. Functional canonicalization was narrowed to that bookkeeping field; exact weight equality remained required.
- The earlier
neuronaut-confirm-1023iworker released itself when changing artifact files stalled the mirror. Its dictionary log reached 1,408 SDXS features; only the first 128-feature chunk (three seeds, 384 images) is retained. This is a pipeline failure, not a completed dictionary. - Full dictionaries, semantic labels, native exemplar indexing, independent semantic/causal evaluation, and Space integration remain unfinished. Prepared scripts do not establish executed results.
Provenance and navigation
runs/neuronaut-usae-1023b: original K10 experiment and activation shards.
runs/neuronaut-capture-1023g: capacity study; complete aligned K32/K128/K256 and independent K32/K128 checkpoints.
runs/neuronaut-confirm-1023h: exact replay and diagnosed bookkeeping-check failure.
runs/neuronaut-confirm-1023i: selected bit-exact replays, fresh confirmation, native preservation controls and retained partial feature-only dictionary.
Pinned source revisions, exact executed source, frozen inputs, configs, training traces, hashes, and release records are included. CURRENT.md tracks the present state. Source archives are historical snapshots; current host uploader code is separate and keeps credentials on the host.