Unlike the earlier fixed-3-image adapters, this v2 reads any number of screenshots from 1 to 6 at inference time, at a fixed 2x pixel compression per image.

Performance (GPT-4.1 LLM-judge on 500 test examples)

k (images per sample)	base Qwen3-VL-4B @ 0x (no compression, no SFT)	This adapter @ 2x	Δ
1	0.958	0.946	-0.012
2	0.912	0.950	+0.038
3	0.892	0.954	+0.062
4	0.856	0.936	+0.080

Base@0x degrades sharply as the number of images grows (distractor confusion: 0.958 → 0.856 from k=1 to k=4). The 2x adapter is nearly k-invariant: 0.946 → 0.936 from k=1 to k=4 — only 0.010 drop.

At 2x compression, this adapter exceeds the uncompressed no-SFT baseline at k≥2 and nearly matches it at k=1 (0.946 vs 0.958), demonstrating that reader SFT fully compensates for both 2x compression and multi-image distractor confusion.

Training setup

Method: LoRA (r=256, alpha=256, targets=all linear layers including ViT; freeze_vision_tower=false)
Base model: Qwen/Qwen3-VL-4B-Instruct
Framework: LLaMA-Factory (fork)
Optimizer: cosine LR, peak 1e-5, warmup 3%
Effective batch: 32 (per_device=1 × 8 GPUs × grad_accum=4)
cutoff_len: 10240 (scaled per compression — larger for smaller compression where each image takes more tokens)
Hardware: 8× H100 80GB, DeepSpeed ZeRO-2, bf16
Training data: mixed top3 (fixed k=3, 104k) + vark (k=1..6, 104k) at 1:1 ratio, 2 epochs over 208k samples (final adapter at step 13004).

Data — variable-k retrieval-augmented multi-image

Built from the Chrisyichuan/screenshot-training-natural-filtered-v2 QA dataset (~104k train examples):

For each query, retrieve top-6 screenshots from a Qwen3-VL-2B embedding index (dora-ls005 checkpoint) over 28M Wikipedia tiles.
Per-sample, uniformly sample k ∈ {1, 2, 3, 4, 5, 6}.
Compose the k-image set: gold always included + (k-1) non-gold hits (padded with gold repeats if retrieval has fewer than k-1 non-gold hits).
Randomize gold position among the k images (prevents positional shortcuts).
Apply 2x compression (each dimension scaled by 1/sqrt(2) via PIL LANCZOS).
Train the reader with <image>×k \n {query} → gold answer.

Distribution: ~17k train samples per k value (uniform). eval/test sets likewise stratified over k∈{1..6}.

Gold-retrieval rate at top-6 across splits: ~75%. When gold is present at rank 1..6 among the retrieved hits, composition uses only retrieval results (else gold is always still included by construction).

Usage

python

from transformers import AutoProcessor, Qwen3VLForConditionalGeneration
from peft import PeftModel
import torch

base = "Qwen/Qwen3-VL-4B-Instruct"
adapter = "Chrisyichuan/qwen3vl-4b-wiki-screenshot-multik-2x-lora"

model = Qwen3VLForConditionalGeneration.from_pretrained(base, torch_dtype=torch.bfloat16).cuda()
model = PeftModel.from_pretrained(model, adapter).merge_and_unload()
processor = AutoProcessor.from_pretrained(base)

# k can be any integer in 1..6. Images must already be 2x-compressed.
messages = [{"role": "user", "content": [
    {"type": "image", "image": img_1},
    # ... up to img_6 ...
    {"type": "text",  "text": your_question},
]}]
# ... standard Qwen3-VL inference

Notes / limitations

Pixel budget is fixed at 2x. If your deployment can afford less compression, use the 2x sibling; if more, use the 4x sibling. Mixing compression at inference is untested.
Training always included gold in the image set. If your retriever misses the gold at inference, this adapter has not seen that distribution — expect degradation on those queries.
For k ∈ {1, 2, 3, 4} this adapter was evaluated with GPT-4.1 LLM-judge. k=5/6 were trained on but not explicitly benchmarked.
v1 of this adapter (trained on fixed k=3 only) achieved slightly higher k=3 score but collapses at other k. This v2 trades ~0.02 at k=3 for flexibility across k=1..6.
Sister adapters at other compression levels: Chrisyichuan/qwen3vl-4b-wiki-screenshot-multik-3x-lora, Chrisyichuan/qwen3vl-4b-wiki-screenshot-multik-4x-lora.