ComfyUI-Prompt-Calibrator/docs/METHODOLOGY.md

# Local Prompt Calibrator — Methodology

> Goal: a **fully local** ComfyUI feedback loop where a vision‑language model (VLM)
> scores how close a *generated* image is to a *reference* image, and that score +
> a structured difference analysis is used to **calibrate the prompt‑generation
> method** ([ComfyUI‑Prompt‑Builder](../../ComfyUI-Prompt-Builder), the "SxCP" nodes)
> until the generated image matches the reference.

---

## 1. The loop at a glance

```
        ┌──────────────────────────────────────────────┐
        │  REFERENCE image  (the target look)           │
        └───────────────┬──────────────────────────────┘
                        │
   ┌────────────────────▼────────────────┐   calibration deltas
   │ Prompt-Builder (SxCP)  ── "method"  │◄──── (axis nudges / knob
   │  seeded pools + profile knobs       │       overrides / seed move)
   └────────────────────┬────────────────┘
                        │ prompt + negative
   ┌────────────────────▼────────────────┐
   │ T2I model  (SDXL / Flux / Krea2)    │   ← fix the sampler seed while
   └────────────────────┬────────────────┘     searching the prompt axes
                        │ generated image
   ┌────────────────────▼──────────────────────────────────┐
   │ Qwen3-VL JUDGE node  ── the "vllm node"                │
   │  in : reference + generated                            │
   │  out: overall_score 0..1                               │
   │       per-axis {score, ref, gen} over ~20 axes         │
   │         (identity, body, wardrobe, action, affect,     │
   │          camera, render) — target vs current values    │
   │       (local model observes only; no fixes suggested)  │
   └────────────────────┬──────────────────────────────────┘
                        │ score + ref/gen per axis
   ┌────────────────────▼────────────────┐
   │ CALIBRATOR / controller             │
   │  - accumulate per-axis scores        │
   │  - map diffs → axis adjustments      │
   │  - update Prompt-Builder knobs        │
   │  - stop when overall_score ≥ target   │
   │    or max iterations reached          │
   └──────────────────────────────────────┘
```

The novel piece is the **Judge node**. Off‑the‑shelf Qwen‑VL nodes emit free text;
a calibrator needs a **machine‑readable score + per‑axis diffs** so the controller
can act on them. That is what `nodes/qwen_judge.py` in this repo provides.

---

## 2. The VLLM node — what to reuse

You already have the model converted locally:

```
/media/p5/qwen3vl_4b_abliterated_comfy_convert/
  ├── hf_bf16/   ← huihui-ai Qwen3-VL-4B-Instruct **abliterated** (uncensored), bf16
  └── hf_fp8/    ← same model, FP8 (≈4–5 GB, trivially fits the RTX 5090 32 GB)
```

The **abliterated** variant matters: stock Qwen3‑VL will often refuse to "describe or
analyze" adult imagery, which would break the loop. huihui‑ai removed the text‑side
refusal direction, so it scores NSFW reference/generated pairs without bailing.

### Reusable ComfyUI nodes (pick one as the plumbing base)

| Repo | Backend | Multi‑image | Local path | Notes |
|---|---|---|---|---|
| **[hardik-uppal/ComfyUI-QwenVL-MultiImage](https://github.com/hardik-uppal/ComfyUI-QwenVL-MultiImage)** | transformers | ✅ `images` + `images_batch_2/3` | needs tiny tweak | **Best base** — built for "compare these images, describe the differences"; supports FP16 / 8‑bit / 4‑bit **and pre‑quantized FP8** (matches your `hf_fp8`). |
| [IuvenisSapiens/ComfyUI_Qwen3-VL-Instruct](https://github.com/IuvenisSapiens/ComfyUI_Qwen3-VL-Instruct) | transformers | ✅ multi‑image query | HF download | Clean native Qwen3‑VL‑Instruct integration. |
| [jren712/ComfyUI-QwenVL-abliterated](https://github.com/jren712/ComfyUI-QwenVL-abliterated) | transformers | ✅ | abliterated‑oriented | Fork tuned for the abliterated weights. |
| [1038lab/ComfyUI-QwenVL](https://github.com/1038lab/ComfyUI-QwenVL) | **GGUF** (llama.cpp) | ✅ | local GGUF | Use only if you want GGUF; bf16 4B on 32 GB doesn't need it. |

**Recommendation:** don't run any of them *as‑is* for the loop — they only output text.
Instead reuse their **model‑load + `apply_chat_template` + `generate`** plumbing inside
a purpose‑built **Judge node** (this repo) that forces structured JSON output. The
`ComfyUI-QwenVL-MultiImage` loader is the closest template (it already handles two
image batches + FP8).

### Model sizing on 32 GB (RTX 5090) — abliterated, latest Qwen VL

As of June 2026 the **latest Qwen VL family is Qwen3‑VL** (Qwen3.5‑VL shipped early
2026, but abliterated builds of it are **text‑only so far** — no uncensored
Qwen3.5‑*VL* yet). So "latest + uncensored + fits 32 GB" = **Qwen3‑VL‑30B‑A3B abliterated**.
All rows below are huihui‑ai abliterated (uncensored) weights:

| Model (abliterated) | Best precision on 32 GB | ~VRAM | Verdict |
|---|---|---|---|
| **Qwen3‑VL‑30B‑A3B‑Instruct** ([HF](https://huggingface.co/huihui-ai/Huihui-Qwen3-VL-30B-A3B-Instruct-abliterated)) | **nf4 (4‑bit)** or GGUF Q4_K_M | ~18 GB | **Best judge that fits.** MoE → only 3B active, so it's fast despite 30B total. transformers class `Qwen3VLMoeForConditionalGeneration` (auto‑detected by the node). |
| Qwen3‑VL‑8B‑Instruct ([HF](https://huggingface.co/huihui-ai)) | bf16 | ~17 GB | Easy middle ground, no quantization. Clearly better than 4B; drop‑in for the judge node. |
| Qwen3‑VL‑4B‑Instruct (already local) | fp8 / bf16 | ~5 / ~9 GB | Lightweight fallback / fast iteration. |

**Gemma alternative:** Gemma‑3‑27B‑it (abliterated, 4‑bit ~16 GB) is a solid different
visual prior if you want a second opinion, but the Krea2 text encoder + Prompt‑Builder
are already Qwen‑aligned, so staying on Qwen3‑VL keeps the vocabulary consistent.

Download an upgrade and point the node's `model_path` at it:
```bash
hf download huihui-ai/Huihui-Qwen3-VL-30B-A3B-Instruct-abliterated \
  --local-dir /media/p5/models/Qwen3-VL-30B-A3B-abliterated
# then in the Judge node: model_path=<that dir>, precision=nf4
```

Practical note: at nf4 the 30B judge (~18 GB) and an SDXL/Flux T2I model can't always
co‑reside — run them as **separate queue steps** and let ComfyUI unload between; the loop
is sequential anyway. The 8B bf16 judge co‑resides more easily.

---

## 3. Scoring rubric (what the VLM actually returns)

The judge prompts Qwen3‑VL to return **strict JSON** with one overall score and, per axis,
the **target value (`ref`), the current value (`gen`), and the gap (`score`)** — exactly
the *target / current / distance* an agent needs to calibrate. The local model only
observes; it suggests no fixes (a stronger external model owns correction).

```json
{
  "overall_score": 0.0,
  "axes": {
    "subject_count": {"score": 1.0, "ref": "1 woman", "gen": "1 woman"},
    "position":      {"score": 0.3, "ref": "doggy style", "gen": "missionary"},
    "clothing_state":{"score": 0.4, "ref": "red lace lingerie", "gen": "nude"},
    "scene":         {"score": 0.5, "ref": "dim bedroom", "gen": "outdoor"},
    "framing":       {"score": 0.6, "ref": "full body", "gen": "close-up"},
    "lighting_color":{"score": 0.5, "ref": "warm low-key", "gen": "flat daylight"}
  }
}
```

The axis list is **configurable** on the node. The default ~20 axes are grouped as
identity / body / wardrobe / action / affect / camera / render, kept granular so the
*action* cluster (`sexual_act`, `position`, `penetration`, `explicitness`, `body_contact`)
stays discriminative for explicit content. The agent steers each low axis's prompt wording
toward its `ref` value. See [CALIBRATION_POLICY.md](CALIBRATION_POLICY.md).

### Reducing VLM‑as‑judge variance (important)

VLM scoring is noisy and biased. Mitigations baked into the node / recommended:

1. **Position‑bias swap** — run the judge twice with reference/generated order swapped and
   average the per‑axis scores (`swap_eval=True`). Cuts the "first image wins" bias.
2. **Low temperature** (0.0–0.3) + a **fixed rubric** in the system prompt → repeatable scores.
3. **Anchored 0–1 rubric** (0 = unrelated, 0.5 = same category/different details, 1 = near‑identical) so scores are comparable across iterations.
4. **Evidence‑first**: ask the model to state the concrete difference *before* the number; reasoning‑then‑score is measurably more reliable than score‑then‑reasoning.
5. **Average over k T2I seeds** for the *same* prompt if you want the score to reflect the prompt rather than sampler noise — or, cheaper, **freeze the T2I seed** during the axis search and only vary it once at the end.

---

## 4. The calibrator / controller

> **Chosen design: the controller is an external CLI agent, not an in‑graph node.**
> The agent reads the Judge's text/JSON analysis, calibrates the prompt, injects it into
> the `CalibratorPromptReceptor` node, and queues ComfyUI via its HTTP API — one
> `prompt_id` per iteration. See **[AGENT_LOOP.md](AGENT_LOOP.md)** and `agent_bridge.py`.
> The options below describe the *policy* the agent can run.

Prompt‑Builder is a **deterministic, seeded, combinatorial** generator (it is *not* an
LLM). So "calibration" = **searching the space of `(seed, profile, per‑axis overrides)`**
to maximize `overall_score`. Three controller options, easiest → strongest:

1. **Greedy per‑axis hill‑climb (start here).**
   Take the lowest‑scoring axis, rewrite that axis's prompt wording toward its `ref`
   (target) value, regenerate, keep the change if `overall_score` improved, else revert.
   Loop until ≥ target or no axis improves. The agent decides the wording (no machine
   fixes). Implementable with the Prompt‑Builder **For‑Loop Start/End + Accumulator** nodes.

2. **Black‑box optimizer over the knob vector.**
   Encode the exposed knobs as a parameter vector and drive it with Optuna / CMA‑ES /
   a simple bandit, objective = `overall_score`. Better for >3–4 interacting axes; needs
   a thin Python controller node that holds state across iterations.

3. **LLM‑in‑the‑loop rewriter.**
   Feed `diff_analysis` to a (local) text LLM that proposes the next knob settings (or,
   if you move to free‑text prompts, rewrites the prompt). Most flexible, least
   reproducible — use the same abliterated Qwen3 text head to keep it local and uncensored.

**Loop hygiene:** fix resolution/sampler/steps across iterations; freeze T2I seed while
searching; stop on `overall_score ≥ target` (e.g. 0.85) **or** `max_iters`; log every
`(knobs, score, diff)` triple so the search is auditable and resumable.

---

## 5. Concrete build order

1. **Judge node** (this repo, `nodes/qwen_judge.py`) — load local Qwen3‑VL‑4B abliterated,
   take ref+gen, output `overall_score (FLOAT)`, `axis_scores (JSON STRING)`,
   `diff_analysis (STRING)`, `raw (STRING)`. ✅ scaffolded.
2. **Wire the loop** in a workflow: Prompt‑Builder → T2I → Judge → Accumulator, using the
   SxCP For‑Loop nodes; route `overall_score` into the loop's stop condition.
3. **Controller node** — start with greedy per‑axis hill‑climb that reads `diff_analysis`
   and emits knob overrides back into Prompt‑Builder's split control nodes.
4. **Tune the judge** — calibrate the rubric on a handful of known ref/gen pairs; enable
   `swap_eval`; pick temperature; decide if you need to step up to 8B/30B‑A3B.

See [README.md](../README.md) for install/usage of the Judge node.