❌ REJECT

CB LoRA v2 — stronger hyperparams to fix v1 identity drift

20260530-cb-lora-v2 · 2026-05-30 · $8.50 · ~50min training + ~2min inference + ~30s judge (in progress)

What this run is

v2 of the CB identity LoRA, kicked off in response to v1's borderline verdict (1-in-6 cells fully drifted to a stranger, which fails the customer Turing bar even though it passed the loose 80% bar).

Same fixture (14 Wikimedia Commons CB photos), same 6-cell inference matrix (3 prompts × 2 seeds 42/7301), same FLUX-dev base model on Replicate via ostris/flux-dev-lora-trainer. Different hyperparams:

	v1	v2
steps	2000	4000
lora_rank	32	64
trigger	TOKR	TOKR
learning_rate	4e-4	4e-4
resolution	1024	1024
Replicate destination	`ikatomi-cb-fixture-lora-v1`	`ikatomi-cb-fixture-lora-v2`
training cost	~$5	~$8 (2x steps)

Verdict reasoning

IN-FLIGHT. The bar for PROCEED on this run is the customer Turing bar (Thor 2026-05-30):

"comparing orig with variant, an amateur like me does not identify the variant as AI generated"

Operationalized in the judge as: - Per-cell Q1 ≥ 0.7 AND Q4 ≥ 0.7 - Aggregate Q1 pass rate ≥ 95% AND Q4 pass rate ≥ 90% - No cell with Q1 < 0.3 (catastrophic identity loss is a deal-breaker even at low frequency — the v1 front-full-S7301 cell would have failed this)

If v2 passes → architectural recommendation: ship LoRA at rank=64, steps=4000. If v2 fails on Q1 → next probe is fixture expansion (more diverse CB photos). If v2 fails on Q4 → next probe is Magnific photorealistic upscale on outputs.

How to continue this work

in the labs channel: training is in flight, no action needed. Cadence updates
every 5 min on the C0B75KPL50U thread. Verdict will land here automatically.

To kick off a manual sibling probe BEFORE the v2 verdict:

in the labs channel: kick off a v3 with a wider fixture (add 10 more diverse
CB photos from Wikimedia, train at the same v2 hyperparams). ~$8 cost. Run in
parallel with v2 so we have both verdicts.

Artifacts (populating as the run completes)

training_state.json — Replicate training_id + version + 6-cell inference results (populated by scripts/train_cb_lora_v2.py)
CB-LORA-{front-full|upper-front|3q-side}-S{42|7301}.png — 6 inference cells
judge_output.json — per-cell + aggregate verdicts (populated by scripts/judge_lora_identity.py)

Inputs

Same 14 CC-licensed Wikimedia Commons photos at inputs/lora-cb-fixture/cb_*.jpg. Zipped to cb-fixture.zip (83MB) for Replicate upload — zip is regenerable, not committed.

Notes / gotchas

The labs server picks this run up automatically — visible at /run/20260530-cb-lora-v2 from the moment this SUMMARY.md exists.
Don't claim the LoRA looks great from your own eyes (per the v1 lesson). Judge verdict is the deliverable; visual impression is not.

Judge verdict (v1-2026-05-30)

REJECT — Q4 (photographic) 50% below 80% — AI-plastic look. Q1 (identity) 83% OK. Option: pipe outputs through Magnific photorealistic upscale and re-judge.
Pass rates: Q1 identity 83% · Q2 pose 17% · Q3 framing 17% · Q4 photographic 50%
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": "2026-05-30T10:34:20.592637+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/20260530-cb-lora-v2/CB-LORA-front-full-S42.png",
      "q1_identity": {
        "pass": true,
        "score": 0.82,
        "reason": "Facial proportions, jaw shape, eye spacing, nose bridge, and lip shape closely match the reference \u2014 recognizable as the same person despite no glasses and different lighting."
      },
      "q2_pose": {
        "pass": false,
        "score": 0.2,
        "reason": "The prompt calls for standing front-facing, but the generated image shows the person crouching/sitting on the ground in a three-quarter pose."
      },
      "q3_framing": {
        "pass": false,
        "score": 0.25,
        "reason": "The prompt requests a full-body shot but the framing is a mid-body crop showing only down to the feet in a compressed crouching position, not the expected head-to-feet standing frame."
      },
      "q4_photographic": {
        "pass": false,
        "score": 0.52,
        "reason": "The skin has a slightly waxy, AI-smoothed quality and the lighting feels artificially even, giving it a typical diffusion-model portrait look rather than a genuine fashion catalog photograph."
      }
    },
    {
      "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/20260530-cb-lora-v2/CB-LORA-front-full-S7301.png",
      "q1_identity": {
        "pass": true,
        "score": 0.75,
        "reason": "The facial structure, jawline, eye shape, and overall proportions are plausibly the same person\u2014blonde bob, similar nose and lip shape\u2014though the generated face looks slightly softer and more generic."
      },
      "q2_pose": {
        "pass": true,
        "score": 0.72,
        "reason": "The subject is standing front-facing as prompted, though there is a slight lean/tilt rather than a fully upright neutral stance."
      },
      "q3_framing": {
        "pass": false,
        "score": 0.35,
        "reason": "The prompt calls for full-body framing showing head to feet, but the generated image is cropped at roughly mid-thigh, missing the lower legs and feet entirely."
      },
      "q4_photographic": {
        "pass": false,
        "score": 0.45,
        "reason": "The background is an outdoor crowd scene rather than a clean studio backdrop, skin has a slightly waxy AI-generated quality, and the overall composition does not match a clean studio catalog photograph."
      }
    },
    {
      "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/20260530-cb-lora-v2/CB-LORA-upper-front-S42.png",
      "q1_identity": {
        "pass": true,
        "score": 0.82,
        "reason": "Facial proportions, jawline, nose bridge, lip shape, and eye spacing are highly consistent with the reference person, making her recognizable despite slightly smoother skin rendering."
      },
      "q2_pose": {
        "pass": false,
        "score": 0.45,
        "reason": "Prompt requested front-facing but the subject is turned at a three-quarter angle with hand raised to chin, not a neutral front-facing pose."
      },
      "q3_framing": {
        "pass": true,
        "score": 0.72,
        "reason": "Upper-body framing is approximately correct, showing from roughly waist up, though the prompt specified a studio backdrop and this shows an outdoor waterfront background."
      },
      "q4_photographic": {
        "pass": true,
        "score": 0.71,
        "reason": "The image has a mostly photographic quality with natural outdoor lighting, though the skin has a slightly AI-smoothed texture and the garment color-block pattern looks slightly synthetic."
      }
    },
    {
      "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/20260530-cb-lora-v2/CB-LORA-upper-front-S7301.png",
      "q1_identity": {
        "pass": false,
        "score": 0.2,
        "reason": "Neither of the two women in the generated image convincingly matches the reference person's distinctive facial features\u2014sharp cheekbones, specific nose shape, and eye proportions are not preserved."
      },
      "q2_pose": {
        "pass": false,
        "score": 0.1,
        "reason": "The prompt asked for a front-facing upper-body pose but the generated image shows two full-body figures in an outdoor event setting, not a front-facing studio pose."
      },
      "q3_framing": {
        "pass": false,
        "score": 0.05,
        "reason": "The prompt called for an upper-body crop on a clean studio backdrop, but the image shows two full-body figures at an outdoor red-carpet event with a crowd background."
      },
      "q4_photographic": {
        "pass": true,
        "score": 0.65,
        "reason": "The image has a photographic quality and realistic event atmosphere, but shows two people instead of one and has some AI telltales in the crowd and signage rendering."
      }
    },
    {
      "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/20260530-cb-lora-v2/CB-LORA-3q-side-S42.png",
      "q1_identity": {
        "pass": true,
        "score": 0.78,
        "reason": "The jawline, eye shape, nose bridge, and overall facial proportions are a plausible match to the reference person, with the characteristic sharp cheekbones and refined features recognizable."
      },
      "q2_pose": {
        "pass": false,
        "score": 0.2,
        "reason": "The prompt asked for standing, but the generated image shows the person crouching/seated on the ground with a hand-to-chin pose, not standing."
      },
      "q3_framing": {
        "pass": false,
        "score": 0.3,
        "reason": "The prompt called for a three-quarter side framing but the image is shot from a near-frontal low angle, and the person appears seated rather than standing on a studio backdrop."
      },
      "q4_photographic": {
        "pass": false,
        "score": 0.38,
        "reason": "The image has a noticeably AI-plastic, waxy quality with overly smooth skin, unnaturally rendered clothing geometry around the lower body, and the feet/shoes area looks anatomically distorted and CGI-like."
      }
    },
    {
      "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/20260530-cb-lora-v2/CB-LORA-3q-side-S7301.png",
      "q1_identity": {
        "pass": true,
        "score": 0.72,
        "reason": "Facial proportions, jawline, eye shape, and blonde bob hairstyle are broadly consistent with the reference, though the nose appears slightly softer and the overall face feels somewhat smoothed."
      },
      "q2_pose": {
        "pass": false,
        "score": 0.35,
        "reason": "The subject is standing front-facing toward the camera rather than in a three-quarter side pose as prompted."
      },
      "q3_framing": {
        "pass": false,
        "score": 0.45,
        "reason": "The framing cuts off below the hips rather than showing a proper three-quarter (knee-up) framing, and the background is an outdoor event setting rather than a clean studio backdrop."
      },
      "q4_photographic": {
        "pass": true,
        "score": 0.72,
        "reason": "The image has reasonable photographic qualities with natural-looking skin texture and bokeh background, though the skin tone appears slightly over-tanned and the lighting feels slightly even/AI-processed."
      }
    }
  ],
  "summary": {
    "n_cells": 6,
    "q1_identity_pass_rate": 0.833,
    "q2_pose_pass_rate": 0.167,
    "q3_framing_pass_rate": 0.167,
    "q4_photographic_pass_rate": 0.5,
    "all_axes_pass_rate": 0.0,
    "proceed_threshold": 0.8,
    "verdict": "REJECT",
    "verdict_reason": "Q4 (photographic) 50% below 80% \u2014 AI-plastic look. Q1 (identity) 83% OK. Option: pipe outputs through Magnific photorealistic upscale and re-judge."
  }
}

training_state.json

{
  "zip_url": "https://api.replicate.com/v1/files/NmM4N2IwOWUtZDBiNS00N2UwLWExZDAtMDgxNmRjZjI1MTJj.zip",
  "training_id": "e2gehcjtfnrmw0cyf0patm48nc",
  "training_result": {
    "status": "succeeded",
    "version": "thor-reef/ikatomi-cb-fixture-lora-v2:74854c75c4cfef2349feed9977044e3700d66b63cde77ca59184cfb07749a24a",
    "created_at": "2026-05-30T09:17:13.213Z",
    "started_at": "2026-05-30T09:17:43.398882438Z",
    "completed_at": "2026-05-30T10:29:15.299095308Z"
  },
  "inference": [
    {
      "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",
      "seed": 42,
      "image_path": "/Users/claude/Projects/ikatomi-workspace/ikatomi-labs/runs/20260530-cb-lora-v2/CB-LORA-front-full-S42.png",
      "elapsed_s": 20.3
    },
    {
      "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",
      "seed": 7301,
      "image_path": "/Users/claude/Projects/ikatomi-workspace/ikatomi-labs/runs/20260530-cb-lora-v2/CB-LORA-front-full-S7301.png",
      "elapsed_s": 16.8
    },
    {
      "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",
      "seed": 42,
      "image_path": "/Users/claude/Projects/ikatomi-workspace/ikatomi-labs/runs/20260530-cb-lora-v2/CB-LORA-upper-front-S42.png",
      "elapsed_s": 17.0
    },
    {
      "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",
      "seed": 7301,
      "image_path": "/Users/claude/Projects/ikatomi-workspace/ikatomi-labs/runs/20260530-cb-lora-v2/CB-LORA-upper-front-S7301.png",
      "elapsed_s": 16.8
    },
    {
      "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",
      "seed": 42,
      "image_path": "/Users/claude/Projects/ikatomi-workspace/ikatomi-labs/runs/20260530-cb-lora-v2/CB-LORA-3q-side-S42.png",
      "elapsed_s": 17.0
    },
    {
      "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",
      "seed": 7301,
      "image_path": "/Users/claude/Projects/ikatomi-workspace/ikatomi-labs/runs/20260530-cb-lora-v2/CB-LORA-3q-side-S7301.png",
      "elapsed_s": 16.9
    }
  ]
}