AbstractPhil's picture
Anima adapter is ComfyUI-loadable: the node pack now supports the MiniTrainDIT family (verified live on anima_baseV10, 28 sites, toggle-off bit-exact) — caveat corrected in the catalog and in the file metadata
41865a8 verified
|
Raw
History Blame Contribute Delete
40.5 kB
# AMoE-LoRA adapter catalog
Every trained adapter from the r2 diffusion campaign, converted to safetensors, named for what it does, and documented with the number that was actually measured.
Two views of the same tensors:
| directory | naming | for |
|---|---|---|
| `raw/` | original experiment paths and filenames | reproducing the record |
| `production/` | function-first names + a usage card in the file's own metadata | actually using them |
The tensors are identical between the two; only names and metadata differ. Every file was reloaded and compared bitwise against its source after writing.
- **Format + loader rules:** [SCHEMA.md](SCHEMA.md)
- **Framework:** [amoe-lora](https://github.com/AbstractEyes/amoe-lora) (`pip install git+https://github.com/AbstractEyes/amoe-lora`)
- **ComfyUI nodes:** [comfyui-geolip-amoe-lora](https://github.com/AbstractEyes/comfyui-geolip-amoe-lora)
## Before you pick one
**Match the trunk.** These are per-block adapters for a specific denoiser. An SD1.5-shaped stack will *geometrically* attach to any SD1.5-class UNet, but the measurements only hold on the trunk named in its card. Three trunks appear here:
| trunk | objective | which adapters |
|---|---|---|
| `stable-diffusion-v1-5/stable-diffusion-v1-5` (stock) | epsilon | everything under `sd15/` |
| `AbstractPhil/sd15-flow-lune-json-prompt` ckpt-2500 | rectified flow | `lune-json/` |
| `AbstractPhil/sd15-flow-lune-flux` ckpt-18765 | rectified flow | `lune-flux/` |
| `circlestone-labs/Anima` 2B DiT | flow | `anima/`**non-commercial** |
**Multiband stacks are step-gated.** Their three experts are combined by cosine crossfade windows on the normalized timestep, so they need a loader that sets band windows each sampling step. The ComfyUI nodes and `amoe.diffusion.StepGatedSampler` both do; a naive always-on load does not and will misbehave.
**Site order is not execution order.** See [SCHEMA.md](SCHEMA.md) — the mid block is stored LAST. Loaders must map by site identity and verify the width signature.
## Capability adapters
Attachable, and they do something you can point at. Each row's *evidence* line is the measured result with its seed count — read it before you trust it.
### `production/sd15/sd15-relay-grounding-v1-s0.safetensors`
**SD1.5 Relay — Grounding v1 (seed 0)**
`relay` · 16 sites · 160 tensors · 12.9 MB · objective `eps` · seed 0 · trunk `stable-diffusion-v1-5/stable-diffusion-v1-5`
*What it is.* An always-on 16-site relay for the stock SD1.5 epsilon trunk. Improves how closely a generation tracks its prompt without retraining the model. The default pick if you just want the adapter that works.
*What was measured.* exp006, 2 seeds, matched controls: paired val eps-MSE 0.124788 (frozen) -> 0.121698 (s0) / 0.121709 (s1), -2.5%. A parameter-matched LoRA-r32 reached only 0.122674 AND traded away grounding (round-trip +0.1809 zero-shot -> +0.1596) where the relay gained it (+0.2172 / +0.2206). Toggle bit-exact post-train at both seeds. Gates GREW 0.0474 -> 0.0700 (the trunk opts in).
*Caveats.*
- Trained on the stock SD1.5 eps trunk at 512px; other trunks are untested.
- Gauges are paired val-MSE and CLIP round-trip grounding, not aesthetics.
*Raw twin.* `raw/exp006_sd15core_relay/relay_all16_s0.safetensors` · source `relay_all16_s0.pt` from the [exp006_sd15core_relay](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp006_sd15core_relay) package
### `production/sd15/sd15-relay-grounding-v1-s1.safetensors`
**SD1.5 Relay — Grounding v1 (seed 1)**
`relay` · 16 sites · 160 tensors · 12.9 MB · objective `eps` · seed 1 · trunk `stable-diffusion-v1-5/stable-diffusion-v1-5`
*What it is.* Second-seed twin of the grounding relay. Use it to check that an effect you see is the adapter and not a seed lottery.
*What was measured.* exp006 s1: 0.121709 val (frozen 0.124788), grounding +0.2206. Replicates s0 to 5 decimal places — this pair is the campaign's cleanest replication.
*Caveats.*
- See the s0 card; identical recipe, different init seed.
*Raw twin.* `raw/exp006_sd15core_relay/relay_all16_s1.safetensors` · source `relay_all16_s1.pt` from the [exp006_sd15core_relay](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp006_sd15core_relay) package
### `production/sd15/sd15-multiband-coarse2fine-v1-s0.safetensors`
**SD1.5 Multiband — Coarse-to-Fine v1 (seed 0)**
`multiband3` · 16 sites · 160 tensors · 5.0 MB · objective `eps` · seed 0 · trunk `stable-diffusion-v1-5/stable-diffusion-v1-5`
*What it is.* Three sigma-band experts per site, gated per sampling step: HIGH noise handles coarse structure, MID continuity, LOW fine detail. Lesion a band to see what it was carrying — this is the adapter to play with, not just run.
*What was measured.* exp008, 2 seeds: band lesions are SURGICAL 3/3 — disabling a band damages its own noise range 50-200x more than any other (band0 own 0.005911 vs cross 0.000029). In image space (exp010) the step-gated controller lifted round-trip grounding to +0.2135 vs frozen +0.1247, and the HIGH-band lesion is 14x low-pass dominated — coarse-to-fine confirmed by lesion signature.
*Bands.* 0: fidelity/detail (LOW noise) · 1: continuity/semantics (MID) · 2: diversity/structure (HIGH noise)
*Caveats.*
- On aggregate eps-MSE a matched rank-48 monolith still edges it (0.121736 vs 0.122004, both seeds). The value is the band structure, not the loss number — compare against sd15-monolith-r48-control.
- Band edges (0.35/0.75) and crossfade (0.06) are fixed law constants, not tunables.
*Raw twin.* `raw/exp008_multiband/mb3_s0.safetensors` · source `mb3_s0.pt` from the [exp008_multiband](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp008_multiband) package
### `production/sd15/sd15-multiband-coarse2fine-v1-s1.safetensors`
**SD1.5 Multiband — Coarse-to-Fine v1 (seed 1)**
`multiband3` · 16 sites · 160 tensors · 5.0 MB · objective `eps` · seed 1 · trunk `stable-diffusion-v1-5/stable-diffusion-v1-5`
*What it is.* Second-seed twin of the multiband stack.
*What was measured.* exp008 s1: lesion asymmetries replicate near-exactly (band0 own 0.005958 vs cross 0.000029). s0's marginal own-band win did NOT replicate; the surgical structure did.
*Bands.* 0: fidelity/detail (LOW noise) · 1: continuity/semantics (MID) · 2: diversity/structure (HIGH noise)
*Caveats.*
- See the s0 card.
*Raw twin.* `raw/exp008_multiband/mb3_s1.safetensors` · source `mb3_s1.pt` from the [exp008_multiband](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp008_multiband) package
### `production/sd15/sd15-multiband-realdata-v1-s0.safetensors`
**SD1.5 Multiband — Real-Data v1 (seed 0)**
`multiband3` · 16 sites · 160 tensors · 5.0 MB · objective `eps` · seed 0 · trunk `stable-diffusion-v1-5/stable-diffusion-v1-5`
*What it is.* The same three-band mechanism trained on real photographs with multiple entities (qwen-deepfashion-fused) rather than synthetic scenes. Prefer this one for photographic subjects.
*What was measured.* exp011, 2 seeds: frozen 0.142435 -> 0.135215 (s0) / 0.135245 (s1). Adapters pay 2-3x more on real data than on synthetic rows, and the surgical lesion structure replicates 3/3 at both seeds.
*Bands.* 0: fidelity/detail (LOW noise) · 1: continuity/semantics (MID) · 2: diversity/structure (HIGH noise)
*Caveats.*
- The matched monolith again edges the aggregate (0.134184 vs 0.135215).
- Trained at 512px on fashion/person-centric imagery — that is where it was measured.
*Raw twin.* `raw/exp011_fused_multiband/multiband3_s0.safetensors` · source `multiband3_s0.pt` from the [exp011_fused_multiband](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp011_fused_multiband) package
### `production/sd15/sd15-multiband-realdata-v1-s1.safetensors`
**SD1.5 Multiband — Real-Data v1 (seed 1)**
`multiband3` · 16 sites · 160 tensors · 5.0 MB · objective `eps` · seed 1 · trunk `stable-diffusion-v1-5/stable-diffusion-v1-5`
*What it is.* Second-seed twin of the real-data multiband stack.
*What was measured.* exp011 s1: 0.135245 val, lesions surgical 3/3 — seed-robust.
*Bands.* 0: fidelity/detail (LOW noise) · 1: continuity/semantics (MID) · 2: diversity/structure (HIGH noise)
*Caveats.*
- See the s0 card.
*Raw twin.* `raw/exp011_fused_multiband/multiband3_s1.safetensors` · source `multiband3_s1.pt` from the [exp011_fused_multiband](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp011_fused_multiband) package
### `production/lune-json/lunejson-relay-fidelity-v1-s0.safetensors`
**Lune-JSON Relay — Fidelity v1 (seed 0)**
`relay` · 16 sites · 160 tensors · 12.9 MB · objective `flow` · seed 0 · trunk `AbstractPhil/sd15-flow-lune-json-prompt@checkpoint-00002500`
*What it is.* The first certified diffusion relay, on the SD15-Lune rectified-flow finetune. Belongs on that trunk with flow sampling, not on stock SD1.5.
*What was measured.* exp001: paired val flow-MSE 0.522410 (frozen) -> 0.517960, -0.85%. The parameter-matched LoRA-r32 landed BELOW frozen (0.525033) and degraded round-trip grounding to +0.1694. Toggle bit-exact post-train.
*Caveats.*
- s0 round-trip grounding is +0.2052, which is BELOW the +0.2099 zero-shot wall for this trunk — it does not 'preserve' the wall as first reported; only s1 (+0.2158) clears it. The relay-beats-LoRA ordering is unaffected.
- The LoRA control arm was run at s0 only, so 'relay >= matched LoRA' is 1-seed here (it is 2-seed on exp006).
- Gates SHRANK on this trunk (0.0474 -> 0.0432) where they grew on the stock core — gate dynamics are substrate-dependent.
*Raw twin.* `raw/exp001_sd15_relay/relay_all16_s0.safetensors` · source `relay_all16_s0.pt` from the [exp001_sd15_relay](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp001_sd15_relay) package
### `production/lune-json/lunejson-relay-fidelity-v1-s1.safetensors`
**Lune-JSON Relay — Fidelity v1 (seed 1)**
`relay` · 16 sites · 160 tensors · 12.9 MB · objective `flow` · seed 1 · trunk `AbstractPhil/sd15-flow-lune-json-prompt@checkpoint-00002500`
*What it is.* Second-seed twin of the Lune-JSON relay; the better-grounded of the pair.
*What was measured.* exp001 s1: val 0.517623 vs frozen 0.522410; round-trip +0.2158, which does clear the +0.2099 wall.
*Caveats.*
- No LoRA control was run at this seed.
*Raw twin.* `raw/exp001_sd15_relay/relay_all16_s1.safetensors` · source `relay_all16_s1.pt` from the [exp001_sd15_relay](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp001_sd15_relay) package
### `production/lune-flux/luneflux-blob-structure-lambda1-s0.safetensors`
**Lune-Flux Blob-Structure — lambda 1.0 (seed 0)**
`multiband3` · 16 sites · 160 tensors · 5.0 MB · objective `flow` · seed 0 · trunk `AbstractPhil/sd15-flow-lune-flux@checkpoint-00018765`
*What it is.* Multiband stack whose HIGH-noise band was additionally supervised on foreground structure (segmentation blobs). The demonstration artifact of the conditioning law: structural supervision pays on a flow trunk.
*What was measured.* exp013 dose curve at lambda=1.0: the HIGH-band foreground gauge moves -8.3% relative to the uniform control, with the common gauge +0.4% (inside the 0.5% tolerance). lambda~1 is the declared operating point. The same coupling on an eps trunk moved +0.03%/-1.0% (exp012, 2 seeds) — a ~125-200x effect ratio, because flow recovers x0 = x_t - sigma*v exactly at every noise level while eps divides by a vanishing sqrt(alpha_bar) in exactly the supervised band.
*Bands.* 0: fidelity/detail (LOW noise) · 1: continuity/semantics (MID) · 2: diversity/structure (HIGH noise)
*Caveats.*
- THIS FILE IS THE lambda=1.0 DOSE POINT, not the lambda=0.5 candidate its experiment README describes. The bed saved to a lambda-less filename and each dose run overwrote the previous; the lambda=0.5 s0 weights no longer exist anywhere. The s1 file IS lambda=0.5. Do not compare s0 against s1 as a seed pair.
- Belongs on the base-Lune flux trunk with flow sampling.
*Raw twin.* `raw/exp013_blob_flow/blob_mb3_s0.safetensors` · source `blob_mb3_s0.pt` from the [exp013_blob_flow](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp013_blob_flow) package
### `production/lune-flux/luneflux-blob-structure-lambda05-s1.safetensors`
**Lune-Flux Blob-Structure — lambda 0.5 (seed 1)**
`multiband3` · 16 sites · 160 tensors · 5.0 MB · objective `flow` · seed 1 · trunk `AbstractPhil/sd15-flow-lune-flux@checkpoint-00018765`
*What it is.* The lambda=0.5 arm of the same conditioning-law demonstration — a lighter structural coupling.
*What was measured.* exp013 s1 at lambda=0.5: blob gauge 0.124435 vs uniform control 0.129276, -3.7% relative (s0 at the same lambda was -5.9%); both cleared the preregistered 5x bar over the eps reference.
*Bands.* 0: fidelity/detail (LOW noise) · 1: continuity/semantics (MID) · 2: diversity/structure (HIGH noise)
*Caveats.*
- lambda differs from the s0 file in this family (1.0 vs 0.5) — they are dose points, not a seed pair.
- Belongs on the base-Lune flux trunk with flow sampling.
*Raw twin.* `raw/exp013_blob_flow/blob_mb3_s1.safetensors` · source `blob_mb3_s1.pt` from the [exp013_blob_flow](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp013_blob_flow) package
### `production/anima/anima-dit-relay-v1-epoch4.safetensors`
**Anima 2B DiT Relay v1 (epoch 4) — NON-COMMERCIAL**
`relay` · 28 sites · 280 tensors · 28.6 MB · objective `flow` · seed 0 · trunk `circlestone-labs/Anima@anima-base-v1.0` · **NC (CircleStone NC + NVIDIA Open Model License)**
*What it is.* A 28-site bf16 relay for the Anima 2B DiT, trained end-to-end through the diffusion-pipe fork. Proof the machinery crosses architectures.
*What was measured.* exp004, single seed: training loss 0.1229 -> 0.1154 (improving) while a native LoRA-r16 control on the identical run went 0.1375 -> 0.1424 (degrading) at 2.6x the parameters.
*Caveats.*
- DERIVED FROM NON-COMMERCIAL WEIGHTS — non-commercial use only.
- This is the weakest-evidenced deliverable in the campaign: exp004 produced no results.json, so the comparison rests on training-loss endpoints rather than a paired held-out validation.
- All 28 gate scalars still read exactly -3.0. bf16 resolution there is 0.0078, so any gate movement is bounded below that — the 'trunk opts in' gate reading from the SD1.5 experiments has no counterpart here.
- Needs an Anima checkpoint: it is a 28-site, 2048-wide DiT stack and will not attach to an SD1.5/SDXL UNet (the loader refuses rather than mis-mapping). Verified working in ComfyUI on anima_baseV10.
*Raw twin.* `raw/exp004_anima_relay/aleph_relay.safetensors` · source `aleph_relay.pt` from the [exp004_anima_relay](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp004_anima_relay) package
## Matched controls
The falsifiers and reference arms. Published so you can A/B the claims yourself rather than take them on faith. Several of these BEAT their treatment arm on the aggregate metric — that is stated, not buried.
### `production/controls/sd15-monolith-r48-control-s0.safetensors`
**SD1.5 Monolith r48 — matched control (seed 0)**
`mono` · 16 sites · 64 tensors · 4.8 MB · objective `eps` · seed 0 · trunk `stable-diffusion-v1-5/stable-diffusion-v1-5`
*What it is.* The parameter-matched single-expert falsifier for the multiband stack. Attach it to A/B whether banding buys anything on YOUR gauge.
*What was measured.* exp008: 0.121736 (s0) / 0.121745 (s1) vs multiband 0.122004 / 0.121970 — the monolith WINS on aggregate eps-MSE at both seeds. Architecture alone does not pay; differentiated pressure has to.
*Caveats.*
- A control, published as evidence. It has no band structure to lesion and no coarse-to-fine behaviour.
*Raw twin.* `raw/exp008_multiband/mono48_s0.safetensors` · source `mono48_s0.pt` from the [exp008_multiband](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp008_multiband) package
### `production/controls/sd15-monolith-r48-control-s1.safetensors`
**SD1.5 Monolith r48 — matched control (seed 1)**
`mono` · 16 sites · 64 tensors · 4.8 MB · objective `eps` · seed 1 · trunk `stable-diffusion-v1-5/stable-diffusion-v1-5`
*What it is.* Second-seed twin of the monolith control.
*What was measured.* exp008 s1: 0.121745 vs multiband 0.121970.
*Caveats.*
- See the s0 card.
*Raw twin.* `raw/exp008_multiband/mono48_s1.safetensors` · source `mono48_s1.pt` from the [exp008_multiband](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp008_multiband) package
### `production/controls/sd15-monolith-r48-realdata-control-s0.safetensors`
**SD1.5 Monolith r48 Real-Data — matched control (seed 0)**
`mono` · 16 sites · 64 tensors · 4.8 MB · objective `eps` · seed 0 · trunk `stable-diffusion-v1-5/stable-diffusion-v1-5`
*What it is.* Monolith control for the real-data multiband stack.
*What was measured.* exp011: 0.134184 (s0) / 0.134221 (s1) vs multiband 0.135215 / 0.135245 — monolith wins the aggregate on real data too.
*Caveats.*
- A control, published as evidence.
*Raw twin.* `raw/exp011_fused_multiband/monolith_s0.safetensors` · source `monolith_s0.pt` from the [exp011_fused_multiband](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp011_fused_multiband) package
### `production/controls/sd15-monolith-r48-realdata-control-s1.safetensors`
**SD1.5 Monolith r48 Real-Data — matched control (seed 1)**
`mono` · 16 sites · 64 tensors · 4.8 MB · objective `eps` · seed 1 · trunk `stable-diffusion-v1-5/stable-diffusion-v1-5`
*What it is.* Second-seed twin.
*What was measured.* exp011 s1: 0.134221.
*Caveats.*
- A control, published as evidence.
*Raw twin.* `raw/exp011_fused_multiband/monolith_s1.safetensors` · source `monolith_s1.pt` from the [exp011_fused_multiband](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp011_fused_multiband) package
### `production/controls/sd15-monolith-r32-control-s0.safetensors`
**SD1.5 Monolith r32 — dispatch falsifier (seed 0)**
`mono` · 16 sites · 64 tensors · 3.3 MB · objective `eps` · seed 0 · trunk `stable-diffusion-v1-5/stable-diffusion-v1-5`
*What it is.* The matched monolith that falsified the 4-expert dispatch bank.
*What was measured.* exp007: monolith 0.121777 (s0) / 0.121799 (s1) BEAT the 4-expert dispatch bank 0.122015 / 0.122011 while using 11% FEWER parameters.
*Caveats.*
- A control, published as evidence.
*Raw twin.* `raw/exp007_amoe_lora/mono_core_s0.safetensors` · source `mono_core_s0.pt` from the [exp007_amoe_lora](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp007_amoe_lora) package
### `production/controls/sd15-monolith-r32-control-s1.safetensors`
**SD1.5 Monolith r32 — dispatch falsifier (seed 1)**
`mono` · 16 sites · 64 tensors · 3.3 MB · objective `eps` · seed 1 · trunk `stable-diffusion-v1-5/stable-diffusion-v1-5`
*What it is.* Second-seed twin.
*What was measured.* exp007 s1: 0.121799.
*Caveats.*
- A control, published as evidence.
*Raw twin.* `raw/exp007_amoe_lora/mono_core_s1.safetensors` · source `mono_core_s1.pt` from the [exp007_amoe_lora](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp007_amoe_lora) package
### `production/controls/sd15-blob-on-eps-control-s0.safetensors`
**SD1.5 Blob-on-eps — conditioning-law control (seed 0)**
`multiband3` · 16 sites · 160 tensors · 5.0 MB · objective `eps` · seed 0 · trunk `stable-diffusion-v1-5/stable-diffusion-v1-5`
*What it is.* Blob supervision applied to an EPSILON trunk — the control that shows the conditioning law from its negative side. Compare against the flow-trunk lune-flux blob stacks.
*What was measured.* exp012, 2 seeds: INERT. s0 moved the HIGH-band gauge 0.111495 vs the uniform 0.111528 (+0.03%, a 3.3e-05 margin) and s1 INVERTED to 0.112651. The keeper from this experiment is a different measurement: on the role-aligned foreground gauge multiband 0.111528 beats the matched monolith 0.123388 (frozen 0.155490) by ~10% at both seeds.
*Bands.* 0: fidelity/detail (LOW noise) · 1: continuity/semantics (MID) · 2: diversity/structure (HIGH noise)
*Caveats.*
- The shipped results.json reads "hit": true for the s0 blob gauge. That is a 3.3e-05 margin that inverts at s1 — the honest verdict is INERT. Do not read the raw flag as a pass.
- A control, published as evidence.
*Raw twin.* `raw/exp012_blob_supervised/blob_mb3_s0.safetensors` · source `blob_mb3_s0.pt` from the [exp012_blob_supervised](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp012_blob_supervised) package
### `production/controls/sd15-blob-on-eps-control-s1.safetensors`
**SD1.5 Blob-on-eps — conditioning-law control (seed 1)**
`multiband3` · 16 sites · 160 tensors · 5.0 MB · objective `eps` · seed 1 · trunk `stable-diffusion-v1-5/stable-diffusion-v1-5`
*What it is.* Second-seed twin of the eps blob control.
*What was measured.* exp012 s1: 0.112651 vs uniform 0.111528 — the s0 'hit' inverts here.
*Bands.* 0: fidelity/detail (LOW noise) · 1: continuity/semantics (MID) · 2: diversity/structure (HIGH noise)
*Caveats.*
- A control, published as evidence.
*Raw twin.* `raw/exp012_blob_supervised/blob_mb3_s1.safetensors` · source `blob_mb3_s1.pt` from the [exp012_blob_supervised](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp012_blob_supervised) package
### `production/controls/luneflux-uniform-mb3-control-s0.safetensors`
**Lune-Flux Uniform Multiband — control (seed 0)**
`multiband3` · 16 sites · 160 tensors · 5.0 MB · objective `flow` · seed 0 · trunk `AbstractPhil/sd15-flow-lune-flux@checkpoint-00018765`
*What it is.* The un-blob-supervised multiband stack on the same flow trunk: the reference arm the blob stacks are measured against.
*What was measured.* exp013: the uniform reference against which blob supervision moved -5.9% (lambda 0.5) and -8.3% (lambda 1.0) on the HIGH-band foreground gauge.
*Bands.* 0: fidelity/detail (LOW noise) · 1: continuity/semantics (MID) · 2: diversity/structure (HIGH noise)
*Caveats.*
- Saved by the same lambda-less filename scheme as its treatment arm; this s0 file is from the lambda=1.0 run.
- A control, published as evidence.
*Raw twin.* `raw/exp013_blob_flow/uniform_mb3_s0.safetensors` · source `uniform_mb3_s0.pt` from the [exp013_blob_flow](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp013_blob_flow) package
### `production/controls/luneflux-uniform-mb3-control-s1.safetensors`
**Lune-Flux Uniform Multiband — control (seed 1)**
`multiband3` · 16 sites · 160 tensors · 5.0 MB · objective `flow` · seed 1 · trunk `AbstractPhil/sd15-flow-lune-flux@checkpoint-00018765`
*What it is.* Second-seed twin of the flow uniform control.
*What was measured.* exp013 s1: uniform 0.129276 vs blob 0.124435.
*Bands.* 0: fidelity/detail (LOW noise) · 1: continuity/semantics (MID) · 2: diversity/structure (HIGH noise)
*Caveats.*
- A control, published as evidence.
*Raw twin.* `raw/exp013_blob_flow/uniform_mb3_s1.safetensors` · source `uniform_mb3_s1.pt` from the [exp013_blob_flow](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp013_blob_flow) package
## Negatives and falsified artifacts
Evidence, explicitly not capability. Dispatch banks and the cond adapter are **load-only**`amoe.attach` refuses them by design and tells you why. They are here because a falsification with weights attached is worth more than one without.
### `production/negatives/sd15-multiband-freqroles-inert-s0.safetensors`
**SD1.5 Multiband Frequency-Roles — measured inert (seed 0)**
`multiband3` · 16 sites · 160 tensors · 5.0 MB · objective `eps` · seed 0 · trunk `stable-diffusion-v1-5/stable-diffusion-v1-5`
*What it is.* Multiband trained with frequency-reweighted role pressure (LOW gets high-pass loss, HIGH gets low-pass). Kept as evidence that this particular pressure is not the one that pays.
*What was measured.* exp009, 2 seeds + rejudge: role gauges improved over uniform multiband in both gauges at both seeds (4/4 directional) but by 0.05-0.2% — nearly collinear with the base objective. In image space (exp010) its lesion signature is IDENTICAL to uniform's (band2 LP 0.0724 vs 0.0727) and it slightly cost grounding (+0.1996 vs +0.2135).
*Bands.* 0: fidelity/detail (LOW noise) · 1: continuity/semantics (MID) · 2: diversity/structure (HIGH noise)
*Caveats.*
- Published as a negative. Band roles need qualitatively different supervision, not frequency reweighting.
- The as-run reference arms lacked role gauges; that instrument gap was confessed and rejudged.
*Raw twin.* `raw/exp009_bandroles/roles_s0.safetensors` · source `roles_s0.pt` from the [exp009_bandroles](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp009_bandroles) package
### `production/negatives/sd15-multiband-freqroles-inert-s1.safetensors`
**SD1.5 Multiband Frequency-Roles — measured inert (seed 1)**
`multiband3` · 16 sites · 160 tensors · 5.0 MB · objective `eps` · seed 1 · trunk `stable-diffusion-v1-5/stable-diffusion-v1-5`
*What it is.* Second-seed twin.
*What was measured.* exp009 s1: 0.121971 vs uniform 0.122004.
*Bands.* 0: fidelity/detail (LOW noise) · 1: continuity/semantics (MID) · 2: diversity/structure (HIGH noise)
*Caveats.*
- Published as a negative.
*Raw twin.* `raw/exp009_bandroles/roles_s1.safetensors` · source `roles_s1.pt` from the [exp009_bandroles](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp009_bandroles) package
### `production/negatives/sd15-dispatch-bank-a4-falsified-s0.safetensors`
**SD1.5 Dispatch Bank A=4 — falsified (seed 0)**
`bank` · 16 sites · 256 tensors · 3.6 MB · objective `eps` · seed 0 · trunk `stable-diffusion-v1-5/stable-diffusion-v1-5`
*What it is.* Four rank-8 experts with a dense signed aleph dispatch keyed on hidden state + sigma. The first MoE-structure test on diffusion. NOT attachable — kept as evidence.
*What was measured.* exp007, 2 seeds: the bank ties/loses to a rank-32 monolith (0.122015 vs 0.121777) with 11% MORE parameters, and sigma-band usage is FLAT (spread 0.0009) with all four experts alive. Aggregation alone creates no specialization.
*Caveats.*
- Load-only: amoe.attach refuses dispatch banks by design.
*Raw twin.* `raw/exp007_amoe_lora/amoe_core_s0.safetensors` · source `amoe_core_s0.pt` from the [exp007_amoe_lora](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp007_amoe_lora) package
### `production/negatives/sd15-dispatch-bank-a4-falsified-s1.safetensors`
**SD1.5 Dispatch Bank A=4 — falsified (seed 1)**
`bank` · 16 sites · 256 tensors · 3.6 MB · objective `eps` · seed 1 · trunk `stable-diffusion-v1-5/stable-diffusion-v1-5`
*What it is.* Second-seed twin.
*What was measured.* exp007 s1: 0.122011, usage spread 0.00385.
*Caveats.*
- Load-only.
*Raw twin.* `raw/exp007_amoe_lora/amoe_core_s1.safetensors` · source `amoe_core_s1.pt` from the [exp007_amoe_lora](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp007_amoe_lora) package
### `production/negatives/luneflux-router-sigmakey-control-s0.safetensors`
**Lune-Flux Router — state+sigma key control (seed 0)**
`bank` · 16 sites · 256 tensors · 3.6 MB · objective `flow` · seed 0 · trunk `AbstractPhil/sd15-flow-lune-flux@checkpoint-00018765`
*What it is.* The control arm of the router experiment: dispatch keyed on hidden state and sigma alone, no text. It won.
*What was measured.* exp014, 2 seeds: 0.573713 (s0) / 0.572191 (s1) — the BEST of the three router arms. Its usage variance by prompt is 2.36e-02 with zero text input, meaning the hidden-state key already routes by prompt on its own.
*Caveats.*
- Load-only: dispatch banks are not attachable.
*Raw twin.* `raw/exp014_te_dispatch/sigma_key_s0.safetensors` · source `sigma_key_s0.pt` from the [exp014_te_dispatch](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp014_te_dispatch) package
### `production/negatives/luneflux-router-sigmakey-control-s1.safetensors`
**Lune-Flux Router — state+sigma key control (seed 1)**
`bank` · 16 sites · 256 tensors · 3.6 MB · objective `flow` · seed 1 · trunk `AbstractPhil/sd15-flow-lune-flux@checkpoint-00018765`
*What it is.* Second-seed twin.
*What was measured.* exp014 s1: 0.572191, the best arm again.
*Caveats.*
- Load-only.
*Raw twin.* `raw/exp014_te_dispatch/sigma_key_s1.safetensors` · source `sigma_key_s1.pt` from the [exp014_te_dispatch](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp014_te_dispatch) package
### `production/negatives/luneflux-router-addrkey-falsified-s0.safetensors`
**Lune-Flux Router — flat address key, falsified (seed 0)**
`bank` · 16 sites · 272 tensors · 4.7 MB · objective `flow` · seed 0 · trunk `AbstractPhil/sd15-flow-lune-flux@checkpoint-00018765`
*What it is.* Dispatch keyed on the raw flattened [32,128] aleph address.
*What was measured.* exp014, 2 seeds: WORST arm (0.577551 / 0.577694) and it KILLS routing — usage variance collapses to 1.40e-05 against the control's 2.36e-02. The high-dimensional disease geometry of the language line, reproduced on a dispatch key.
*Caveats.*
- Load-only.
*Raw twin.* `raw/exp014_te_dispatch/addr_key_s0.safetensors` · source `addr_key_s0.pt` from the [exp014_te_dispatch](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp014_te_dispatch) package
### `production/negatives/luneflux-router-addrkey-falsified-s1.safetensors`
**Lune-Flux Router — flat address key, falsified (seed 1)**
`bank` · 16 sites · 272 tensors · 4.7 MB · objective `flow` · seed 1 · trunk `AbstractPhil/sd15-flow-lune-flux@checkpoint-00018765`
*What it is.* Second-seed twin.
*What was measured.* exp014 s1: 0.577694, routing dead at both seeds.
*Caveats.*
- Load-only.
*Raw twin.* `raw/exp014_te_dispatch/addr_key_s1.safetensors` · source `addr_key_s1.pt` from the [exp014_te_dispatch](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp014_te_dispatch) package
### `production/negatives/luneflux-router-tekey-falsified-s0.safetensors`
**Lune-Flux Router — trained text-encoder key, falsified (seed 0)**
`bank` · 16 sites · 272 tensors · 3.8 MB · objective `flow` · seed 0 · trunk `AbstractPhil/sd15-flow-lune-flux@checkpoint-00018765`
*What it is.* The 'router solidifier': a trained CLIP-pooler adapter feeding the dispatch key.
*What was measured.* exp014, 2 seeds: 0.574217 / 0.574678 — does not beat the state+sigma control (0.573713 / 0.572191). At s0 its by-prompt usage variance (2.598e-02) is actually BELOW its own shuffled null (2.707e-02).
*Caveats.*
- Load-only.
- The bed confessed its own null was wrong (a shuffled-key null measures diversity, not correctness); exp015 rebuilt it as a repeated-key null.
*Raw twin.* `raw/exp014_te_dispatch/te_key_s0.safetensors` · source `te_key_s0.pt` from the [exp014_te_dispatch](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp014_te_dispatch) package
### `production/negatives/luneflux-router-tekey-falsified-s1.safetensors`
**Lune-Flux Router — trained text-encoder key, falsified (seed 1)**
`bank` · 16 sites · 272 tensors · 3.8 MB · objective `flow` · seed 1 · trunk `AbstractPhil/sd15-flow-lune-flux@checkpoint-00018765`
*What it is.* Second-seed twin.
*What was measured.* exp014 s1: 0.574678.
*Caveats.*
- Load-only.
*Raw twin.* `raw/exp014_te_dispatch/te_key_s1.safetensors` · source `te_key_s1.pt` from the [exp014_te_dispatch](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp014_te_dispatch) package
### `production/negatives/luneflux-routerv2-statesigma-control-s0.safetensors`
**Lune-Flux Router v2 — state+sigma control (seed 0)**
`bank` · 16 sites · 256 tensors · 3.6 MB · objective `flow` · seed 0 · trunk `AbstractPhil/sd15-flow-lune-flux@checkpoint-00018765`
*What it is.* Control arm of the corrected router test.
*What was measured.* exp015, 2 seeds: 0.573098 / 0.571703 — again the better arm.
*Caveats.*
- Load-only.
*Raw twin.* `raw/exp015_router_v2/state_sigma_s0.safetensors` · source `state_sigma_s0.pt` from the [exp015_router_v2](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp015_router_v2) package
### `production/negatives/luneflux-routerv2-statesigma-control-s1.safetensors`
**Lune-Flux Router v2 — state+sigma control (seed 1)**
`bank` · 16 sites · 256 tensors · 3.6 MB · objective `flow` · seed 1 · trunk `AbstractPhil/sd15-flow-lune-flux@checkpoint-00018765`
*What it is.* Second-seed twin.
*What was measured.* exp015 s1: 0.571703.
*Caveats.*
- Load-only.
*Raw twin.* `raw/exp015_router_v2/state_sigma_s1.safetensors` · source `state_sigma_s1.pt` from the [exp015_router_v2](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp015_router_v2) package
### `production/negatives/luneflux-routerv2-mhataddr-falsified-s0.safetensors`
**Lune-Flux Router v2 — M-hat address key, FALSIFIED (seed 0)**
`bank` · 16 sites · 288 tensors · 3.7 MB · objective `flow` · seed 0 · trunk `AbstractPhil/sd15-flow-lune-flux@checkpoint-00018765`
*What it is.* The address read through 32 slots and projected 128->4 before routing — the form the disease-geometry lesson demanded. This is where address-as-key closed.
*What was measured.* exp015, 2 seeds with corrected instruments: routing excess over a REPEATED-key null is +2.53e-06 at s0 and -4.25e-06 at s1; match advantage ~0. Val 0.574912 vs the control's 0.573098. Address-as-key in its frozen form is FALSIFIED.
*Caveats.*
- Load-only.
- The open form is TE+AMoE joint training, which this does not test.
*Raw twin.* `raw/exp015_router_v2/mhat_addr_s0.safetensors` · source `mhat_addr_s0.pt` from the [exp015_router_v2](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp015_router_v2) package
### `production/negatives/luneflux-routerv2-mhataddr-falsified-s1.safetensors`
**Lune-Flux Router v2 — M-hat address key, FALSIFIED (seed 1)**
`bank` · 16 sites · 288 tensors · 3.7 MB · objective `flow` · seed 1 · trunk `AbstractPhil/sd15-flow-lune-flux@checkpoint-00018765`
*What it is.* Second-seed twin.
*What was measured.* exp015 s1: 0.573888; routing excess -4.25e-06 (negative).
*Caveats.*
- Load-only.
*Raw twin.* `raw/exp015_router_v2/mhat_addr_s1.safetensors` · source `mhat_addr_s1.pt` from the [exp015_router_v2](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp015_router_v2) package
### `production/negatives/lunejson-addrcond-guidepost-negative-s0.safetensors`
**Lune-JSON Address-Conditioning — guidepost negative (seed 0)**
`cond` · 1 sites · 4 tensors · 0.5 MB · objective `flow` · seed 0 · trunk `AbstractPhil/sd15-flow-lune-json-prompt@checkpoint-00002500`
*What it is.* Injects the frozen byte-trigram aleph address as 32 extra conditioning positions. A conditioning-stream adapter, not a block adapter.
*What was measured.* exp002: appended beside FULL text conditioning it is marginally inert — real vs deranged -0.0009, real vs no-append -0.0068. But the address ALONE (text zeroed) steers: +0.0287 after 2k steps, and the trunk grew its gate 0.047 -> 0.0597 to read it. Redundant-in-context, not dead.
*Caveats.*
- Load-only: a cond-stream adapter, not attachable by the block-adapter verbs.
- The redesign target is complementarity — inject where text is not.
*Raw twin.* `raw/exp002_sd15_addrcond/addrcond_s0.safetensors` · source `addrcond_s0.pt` from the [exp002_sd15_addrcond](https://huggingface.co/AbstractPhil/geolip-aleph-diffusion/tree/main/exp002_sd15_addrcond) package
## Companion pairs
Which adapter to compare against which, when you want to see the claim rather than read it:
| treatment | companion | the question it answers |
|---|---|---|
| `sd15/sd15-relay-grounding-v1-s0` | `controls/…` — see the LoRA note below | does the aleph relay beat a parameter-matched LoRA? (yes, at both seeds) |
| `sd15/sd15-multiband-coarse2fine-v1-s0` | `controls/sd15-monolith-r48-control-s0` | does splitting into sigma bands buy anything over one bigger expert? (not on aggregate loss — but the bands are surgically separable) |
| `sd15/sd15-multiband-realdata-v1-s0` | `controls/sd15-monolith-r48-realdata-control-s0` | same question on real photographs (same answer, larger effects) |
| `lune-flux/luneflux-blob-structure-lambda1-s0` | `controls/luneflux-uniform-mb3-control-s0` | does structural supervision pay on a flow trunk? (yes, -8.3% on the foreground gauge) |
| `controls/sd15-blob-on-eps-control-s0` | `lune-flux/luneflux-blob-structure-lambda05-s1` | the conditioning law from both sides: inert on eps, paying on flow |
| `negatives/sd15-dispatch-bank-a4-falsified-s0` | `controls/sd15-monolith-r32-control-s0` | does a 4-expert dispatch beat one monolith? (no — it loses with 11% more parameters) |
| `negatives/luneflux-routerv2-mhataddr-falsified-s0` | `negatives/luneflux-routerv2-statesigma-control-s0` | can the aleph address route the dispatch? (no — 2.5e-06 over a repeated-key null) |
| `negatives/sd15-multiband-freqroles-inert-s0` | `sd15/sd15-multiband-coarse2fine-v1-s0` | does frequency-reweighted role pressure specialize the bands? (directionally, negligibly) |
**LoRA controls.** The matched LoRA arms (exp001 s0, exp006 s0+s1, and the Anima LoRA-r16) are PEFT-format adapters, not amoe anchors, so they are not renamed here — they stay in their experiment packages under `lora_r32_s*/` and `lora_ctrl/`. Note that `exp006_sd15core_relay/lora_r32_s{0,1}/` and `exp004_anima_relay/lora_ctrl/*/` are missing their `adapter_config.json` on the hub, so PEFT cannot load them directly from there; exp001's does have one.
## Known record corrections
Found while building this catalog, kept here rather than quietly fixed:
1. **exp013 s0 lambda mismatch.** The bed saved to a lambda-less filename and each dose run overwrote the last, so the shipped s0 weights are the **lambda=1.0** dose point while the package's `results.json` describes lambda=0.5. The lambda=0.5 s0 weights no longer exist. s1 is lambda=0.5. The production names encode lambda; s0 and s1 in that family are **dose points, not a seed pair**.
2. **exp012's `"hit": true`.** A 3.3e-05 margin that inverts at s1. The honest verdict is *inert*, which is what the timeline says; any tool reading the raw JSON flag will print the wrong thing.
3. **exp001 'preserves the wall'.** s0's round-trip grounding is +0.2052 against a +0.2099 zero-shot wall — below it. Only s1 (+0.2158) clears it. The relay-vs-LoRA ordering is unaffected.
4. **exp001's LoRA arm is single-seed**, so 'relay ≥ matched LoRA' is 2-seed on exp006 and 1-seed on exp001.
5. **exp004 has no `results.json`.** Its verdict rests on training-loss endpoints and tfevents, not a paired held-out validation — the weakest-evidenced deliverable in the campaign. Its 28 gates also never moved off −3.0 (bf16 resolution there is 0.0078).
6. **The burndown claimed safetensors companions 'for every adapter stack'.** It covered 21 of them; the dispatch banks, monoliths and the cond adapter were missing. That is closed here — all 38 stacks are converted, and the classifier bug that would have mislabeled banks as multiband stacks is fixed in amoe 0.2.1.