✅ PROCEED

CB LoRA identity validation — 6-cell inference matrix

20260529-cb-lora-v1 · 2026-05-29 · $5.30 · ~35min training + ~2min inference (judge pending)

What this run is

A labs probe of the per-identity LoRA path for the v1.9 portrait → variants arc. Trained a flux-dev LoRA on a 14-photo Cate Blanchett fixture (inputs/lora-cb-fixture/, all Wikimedia Commons CC-licensed), then ran a 6-cell inference matrix (3 prompts × 2 seeds) to test whether the LoRA preserves identity well enough across pose / framing variation to ship.

This run validates the multi-photo upload path, which is the architectural pivot after 6 single-photo embedding-conditioned vendors (FASHN, PhotoMaker, InstantID, PuLID, FLUX-PuLID, IPAFID) all failed identity preservation at 8-56%. See /docs/handover/TAKEOVER for the full architectural backstory.

Verdict reasoning

IN-FLIGHT — inference matrix complete (6/6 cells succeeded), but identity-preservation judge has not run yet. The PROCEED bar is:

Q1 face_is_source_identity ≥ 80% across the 6 cells
Q4 looks_photographic ≥ 80% across the 6 cells

If both pass: multi-photo LoRA path is architecturally proven; next step is the webapp port plan. If Q1 fails: identity drift — options are more training steps, higher lora_rank, more fixture photos. If Q4 fails: AI-plastic skin / softbox lighting — pipe outputs through Magnific photorealistic upscale.

Verdict will be updated once the judge runs.

How to continue this work

in the labs channel: run the identity judge against runs/20260529-cb-lora-v1.
4-axis prompt (Q1 face_is_source_identity, Q2 pose_adherence, Q3 framing,
Q4 looks_photographic) using claude-sonnet-4-6 vision, with cb_4.jpg as the
identity reference. Write judge_output.json, update SUMMARY.md verdict.
Cost ~$0.25.

in the labs channel: re-run the CB LoRA inference matrix from runs/20260529-cb-lora-v1
with seeds [19023, 555, 999] instead of [42, 7301]. Same 3 prompts. Cost ~$0.15.
Use the existing trained version (no re-train).

in the labs channel: extend the CB LoRA matrix with a new prompt "<prompt text>"
at seeds 42 and 7301. Don't re-train. Cost ~$0.10.

To validate a different identity → see /docs/prompts → "Train a LoRA for a different identity".

Artifacts

CB-LORA-front-full-S42.png / S7301.png — full-body fashion catalog, white tank + grey shorts, studio backdrop
CB-LORA-upper-front-S42.png / S7301.png — upper-body fashion catalog, white tank, studio backdrop
CB-LORA-3q-side-S42.png / S7301.png — three-quarter side, neutral grey dress, soft natural lighting
training_state.json — Replicate training ID + zip URL + version ref

Inputs

14 CC-licensed Wikimedia Commons photos of Cate Blanchett (Cannes 2026, Berlinale 2023, Camerimage 2024, TIFF 2024, Venice 2024), curated for variety. Live at inputs/lora-cb-fixture/cb_*.jpg. Curation rules in inputs/lora-cb-fixture/README.md.

Notes / gotchas

Trainer: ostris/flux-dev-lora-trainer:26dce37a… on Replicate
Hyperparams: steps=2000, lora_rank=32, resolution=1024, learning_rate=4e-4, trigger_word=TOKR, autocaption=true
Trained LoRA version: thor-reef/ikatomi-cb-fixture-lora-v1:175548be…f510
Inference: 50 steps, guidance 3.5, lora_scale 1.0, 2:3 aspect, 1024px PNG
cb-fixture.zip (83MB) is NOT committed (over GitHub's 50MB recommendation); regenerate with zip -j cb-fixture.zip cb_*.jpg. Also still hosted on Replicate (URL in training_state.json).
Don't claim the LoRA looks great from your own eyes. Visual judgment burned on 2026-05-29 — "outputs look dramatically better" was wrong; judge then flagged 11/12 fail. Thor reviews the labs server page, not your assessment.

Judge verdict (v1-2026-05-30)

PROCEED — Q1 (identity) 83% ≥ 80% AND Q4 (photographic) 83% ≥ 80%. Multi-photo LoRA path validated for this fixture; next step is the webapp port plan.
Pass rates: Q1 identity 83% · Q2 pose 100% · Q3 framing 100% · Q4 photographic 83%
Bar: Q1 ≥ 80% AND Q4 ≥ 80%
Per-cell verdicts in judge_output.json.

Images (6)

Data (2)

judge_output.json

{
  "judge_version": "v1-2026-05-30",
  "model": "claude-sonnet-4-6",
  "judged_at": "2026-05-30T09:00:18.342587+00:00",
  "reference": "inputs/lora-cb-fixture/cb_4.jpg",
  "cells": [
    {
      "label": "CB-LORA-front-full-S42",
      "prompt_id": "front-full",
      "prompt": "a full-body fashion catalog photograph of TOKR, standing front-facing on a clean studio backdrop, wearing a white tank top and grey shorts, natural studio lighting, sharp focus, photographic, no makeup, neutral expression",
      "image_path": "runs/20260529-cb-lora-v1/CB-LORA-front-full-S42.png",
      "q1_identity": {
        "pass": true,
        "score": 0.72,
        "reason": "Blonde bob, sharp facial features, similar angular nose and jawline are recognizable, though the eyes appear slightly wider-set and the overall face is a touch more generic than the reference."
      },
      "q2_pose": {
        "pass": true,
        "score": 0.92,
        "reason": "Subject is clearly standing front-facing with hands in pockets, matching the requested pose well."
      },
      "q3_framing": {
        "pass": true,
        "score": 0.95,
        "reason": "Full-body shot showing head to feet on a clean studio backdrop, exactly as requested."
      },
      "q4_photographic": {
        "pass": true,
        "score": 0.78,
        "reason": "Image reads as a plausible catalog photograph with natural studio lighting, though the skin has a slightly smooth AI-rendered quality and the shadow on the backdrop feels digitally generated."
      }
    },
    {
      "label": "CB-LORA-front-full-S7301",
      "prompt_id": "front-full",
      "prompt": "a full-body fashion catalog photograph of TOKR, standing front-facing on a clean studio backdrop, wearing a white tank top and grey shorts, natural studio lighting, sharp focus, photographic, no makeup, neutral expression",
      "image_path": "runs/20260529-cb-lora-v1/CB-LORA-front-full-S7301.png",
      "q1_identity": {
        "pass": false,
        "score": 0.0,
        "reason": "The generated image shows a young dark-haired male, bearing no resemblance to the blonde middle-aged woman in the reference photo."
      },
      "q2_pose": {
        "pass": true,
        "score": 0.95,
        "reason": "The subject is standing upright, front-facing, which matches the prompt's requirement."
      },
      "q3_framing": {
        "pass": true,
        "score": 0.95,
        "reason": "The image is a full-body shot showing head to feet on a clean studio backdrop, matching the prompt exactly."
      },
      "q4_photographic": {
        "pass": true,
        "score": 0.78,
        "reason": "The image looks like a plausible catalog photograph with natural studio lighting, though the skin has a slightly AI-smooth quality."
      }
    },
    {
      "label": "CB-LORA-upper-front-S42",
      "prompt_id": "upper-front",
      "prompt": "an upper-body fashion catalog photograph of TOKR, front-facing on a clean studio backdrop, wearing a white tank top, natural studio lighting, sharp focus, photographic, no makeup, neutral expression",
      "image_path": "runs/20260529-cb-lora-v1/CB-LORA-upper-front-S42.png",
      "q1_identity": {
        "pass": true,
        "score": 0.82,
        "reason": "The distinctive high cheekbones, wide-set blue-grey eyes, narrow nose bridge, and angular jawline are highly consistent with the reference person."
      },
      "q2_pose": {
        "pass": true,
        "score": 0.92,
        "reason": "The subject is clearly front-facing as prompted, with neutral upright posture."
      },
      "q3_framing": {
        "pass": true,
        "score": 0.88,
        "reason": "Upper-body framing showing roughly mid-torso and above is well-matched to the prompt's request."
      },
      "q4_photographic": {
        "pass": false,
        "score": 0.62,
        "reason": "Skin has a slightly waxy, over-smoothed quality typical of diffusion model portraits, and lighting is too uniformly soft to read as a candid catalog photo."
      }
    },
    {
      "label": "CB-LORA-upper-front-S7301",
      "prompt_id": "upper-front",
      "prompt": "an upper-body fashion catalog photograph of TOKR, front-facing on a clean studio backdrop, wearing a white tank top, natural studio lighting, sharp focus, photographic, no makeup, neutral expression",
      "image_path": "runs/20260529-cb-lora-v1/CB-LORA-upper-front-S7301.png",
      "q1_identity": {
        "pass": true,
        "score": 0.72,
        "reason": "The facial structure, eye shape, nose bridge, and overall proportions are recognizable as the same person despite the hair color change from blonde to dark brown \u2014 the distinctive bone structure and eye set are consistent."
      },
      "q2_pose": {
        "pass": true,
        "score": 0.9,
        "reason": "Subject is front-facing with neutral expression, matching the prompt's front-facing requirement well."
      },
      "q3_framing": {
        "pass": true,
        "score": 0.85,
        "reason": "Upper-body framing is correctly shown from roughly mid-torso up on a clean studio backdrop as requested."
      },
      "q4_photographic": {
        "pass": true,
        "score": 0.78,
        "reason": "The image looks reasonably photographic with natural studio lighting, though skin texture has a slightly smooth AI quality and the background is very uniform, giving a mild diffusion-model look."
      }
    },
    {
      "label": "CB-LORA-3q-side-S42",
      "prompt_id": "3q-side",
      "prompt": "a three-quarter side fashion photograph of TOKR, standing on a clean studio backdrop, looking off-camera, wearing a neutral grey dress, soft natural lighting, sharp focus, photographic, no makeup, neutral expression",
      "image_path": "runs/20260529-cb-lora-v1/CB-LORA-3q-side-S42.png",
      "q1_identity": {
        "pass": true,
        "score": 0.78,
        "reason": "The facial structure\u2014high cheekbones, narrow nose, lip shape, and overall proportions\u2014are consistent with the reference, though the generated face appears slightly smoother and more generic."
      },
      "q2_pose": {
        "pass": true,
        "score": 0.85,
        "reason": "The subject is standing and oriented in a three-quarter side profile, looking off-camera, which matches the prompt well."
      },
      "q3_framing": {
        "pass": true,
        "score": 0.88,
        "reason": "Full three-quarter framing from head to feet is shown against a clean studio backdrop, consistent with a three-quarter side fashion photograph."
      },
      "q4_photographic": {
        "pass": true,
        "score": 0.78,
        "reason": "The image has a believable studio fashion photography aesthetic with natural lighting and sharp focus, though the skin and fabric textures show slight AI-smoothing typical of diffusion models."
      }
    },
    {
      "label": "CB-LORA-3q-side-S7301",
      "prompt_id": "3q-side",
      "prompt": "a three-quarter side fashion photograph of TOKR, standing on a clean studio backdrop, looking off-camera, wearing a neutral grey dress, soft natural lighting, sharp focus, photographic, no makeup, neutral expression",
      "image_path": "runs/20260529-cb-lora-v1/CB-LORA-3q-side-S7301.png",
      "q1_identity": {
        "pass": true,
        "score": 0.72,
        "reason": "The facial structure, high cheekbones, narrow nose, and slim jaw are plausibly consistent with the reference person, though the hair color is redder and the overall rendering is slightly smoothed."
      },
      "q2_pose": {
        "pass": true,
        "score": 0.78,
        "reason": "The figure is standing and looking off-camera at a slight angle, broadly matching a three-quarter side pose as prompted."
      },
      "q3_framing": {
        "pass": true,
        "score": 0.82,
        "reason": "The image shows a full head-to-feet framing in a studio backdrop setting, which is consistent with a three-quarter side fashion photograph."
      },
      "q4_photographic": {
        "pass": true,
        "score": 0.75,
        "reason": "The image has reasonable photographic qualities with soft natural lighting and clean studio backdrop, though the skin has a slightly AI-smooth quality and the overall look is mildly plastic."
      }
    }
  ],
  "summary": {
    "n_cells": 6,
    "q1_identity_pass_rate": 0.833,
    "q2_pose_pass_rate": 1.0,
    "q3_framing_pass_rate": 1.0,
    "q4_photographic_pass_rate": 0.833,
    "all_axes_pass_rate": 0.667,
    "proceed_threshold": 0.8,
    "verdict": "PROCEED",
    "verdict_reason": "Q1 (identity) 83% \u2265 80% AND Q4 (photographic) 83% \u2265 80%. Multi-photo LoRA path validated for this fixture; next step is the webapp port plan."
  }
}

training_state.json

{
  "zip_url": "https://api.replicate.com/v1/files/MjI0Y2FmZDgtOWYzYS00MjkwLTlkZDgtYzI2NWMyMTMyZmUx.zip",
  "training_id": "ey5dzs2z9nrmw0cyej88fasczr",
  "training_result": {
    "status": "succeeded",
    "version": "thor-reef/ikatomi-cb-fixture-lora-v1:175548beef945a51cb30f42da9ce5a5e77d26e0824922e2004efad608a35f510",
    "created_at": "2026-05-29T16:27:59.181Z",
    "started_at": "2026-05-29T16:28:21.288861Z",
    "completed_at": "2026-05-29T17:03:21.448115Z"
  },
  "inference": [
    {
      "label": "CB-LORA-front-full-S42",
      "image_path": "/Users/claude/Projects/ikatomi-workspace/ikatomi-labs/runs/20260529-cb-lora-v1/CB-LORA-front-full-S42.png"
    },
    {
      "label": "CB-LORA-front-full-S7301",
      "image_path": "/Users/claude/Projects/ikatomi-workspace/ikatomi-labs/runs/20260529-cb-lora-v1/CB-LORA-front-full-S7301.png"
    },
    {
      "label": "CB-LORA-upper-front-S42",
      "image_path": "/Users/claude/Projects/ikatomi-workspace/ikatomi-labs/runs/20260529-cb-lora-v1/CB-LORA-upper-front-S42.png"
    },
    {
      "label": "CB-LORA-upper-front-S7301",
      "image_path": "/Users/claude/Projects/ikatomi-workspace/ikatomi-labs/runs/20260529-cb-lora-v1/CB-LORA-upper-front-S7301.png"
    },
    {
      "label": "CB-LORA-3q-side-S42",
      "image_path": "/Users/claude/Projects/ikatomi-workspace/ikatomi-labs/runs/20260529-cb-lora-v1/CB-LORA-3q-side-S42.png"
    },
    {
      "label": "CB-LORA-3q-side-S7301",
      "image_path": "/Users/claude/Projects/ikatomi-workspace/ikatomi-labs/runs/20260529-cb-lora-v1/CB-LORA-3q-side-S7301.png"
    }
  ]
}