# Predator: images, video and music on a 6 GB GPU

Real local-hardware observations from October 20–21, 2026, using a Windows computer with a GTX 1060 6 GB and approximately 28 GiB system RAM. Requests travelled through the real Packet Foundry Warehouse → Registry → Factory relay to a local ComfyUI runtime. These are exploratory deployment and media-generation trials, not a controlled model leaderboard.

## Coverage

The archive retains **143 output files**: **130 images, 11 videos and two music samples**. Of those, 142 were accepted and acknowledged by the Warehouse; one additional Qwen image finished locally but delivery was cancelled after network and lease failures. The image total includes the exact 100-image prompt-complexity suite, ten earlier image outputs, ten Qwen Turbo Q4 follow-up outputs and ten matched Q8 comparison outputs. Two pairs of early accepted images have identical original file hashes, so the archive contains **141 unique original media byte sequences**, while preserving every delivery individually.

Eleven video contact sheets and the image used as LTX video input are derived reference material, not additional model outputs. Two synthetic codec-preflight MP4 fixtures are excluded from the model gallery and generation counts.

Seven submitted jobs ended without media: a cancelled Wan baseline repeat, two LTX storage-guard stops, a Cog attention allocation failure, an H3 legacy-loader failure, an H3 dynamic trial cancelled for impractical runtime, and an incomplete YuE2 CPU probe. Their prompts, available timings, settings and findings appear under `failures` in `dataset.json`. No files are invented for jobs that produced none. Three further installation/schema preflight failures are described separately under `diagnostics`; these were not generation jobs.

## Read the results

`dataset.json` is the authoritative public projection. Each output record links its media and operator-approved workflow, complete input brief, actual conditioning text, seed, settings, available timing and resource measurements, cache observations, technical verification and limited quality findings. Private application, attempt, trace, tenant and host identifiers are omitted. Per-run public workflow hashes and captured source workflow canonical hashes distinguish the privacy-cleaned graph from its original.

The 100-image suite uses ten scene families, five prompt-complexity tiers and adjacent seeds 42/43 for each brief: 50 briefs and exactly 100 images. Briefs range from 16 to 463 whitespace-counted words. Seed and cache state change together between the two arms, and complexity also changes constraint count; the experiment does not isolate prompt length as the sole cause of latency or quality.

The exact ComfyUI and GGUF-node commits, selected model revisions and weight hashes, local conversion identities, recorded Packet Foundry wheel hashes and private integration source hashes are in the dataset. A deployed Packet Foundry Git commit was not recorded in the retained proof, so the current checkout is not substituted for it. No weights are redistributed. Custom workflow node names require the corresponding private integration; the graph alone is not a standalone installer.

## Measurement boundaries

Factory trace duration covers the attempt and differs from native Comfy execution time. Observer node intervals include loading, preparation and event delivery; they are not isolated kernel time or exact weight-loading measurements. Missing timings remain null. Stage telemetry is included where retained. A completed stage in a failed request is not a delivered result.

Quantization labels describe weight storage. Video generators used FP32 arithmetic; an entire pipeline labelled Q8 integer computation would be inaccurate. CUDA parameter checks do not establish physical VRAM residency or absence of Windows WDDM paging. Resource extrema are sampled, often device-wide, and are not continuous process peaks. Cog decoder reserved allocator memory exceeded the card's physical 6 GB.

Cache fields have explicit semantics in the dataset. The second Wan Q8 request followed an approximately 50-minute idle gap: its conditioning was cached, but it is not an immediate warm benchmark. The LTX image-input pair began in a runtime already used for its text pair. Reused encoder probe times describe the earlier encoding, not extra encoding during the cached request. Padded token positions are not content-only token counts, and tokenizer-internal truncation was not independently established.

## Quality and failures remain visible

All 100 complexity-suite thumbnails contained recognizable scenes; exact lettering, fine anatomy and full-resolution trait scores were not systematically assessed. The user listened to both YuE2 files and described them as usable music. Lyrics, intelligibility and comparative music quality were not scored.

Videos include nearly static fox scenes, anatomy and direction errors, two unusable colored-noise clips from the original custom Cog bridge, and corrected Cog outputs with recognizable scenery but poor framing, continuity and requested-action adherence. The original bridge had concrete positional and sampling mismatches. Its noise is not a clean test of Cog quality or Q8 quantization. Five sampled frames and valid decoding do not establish natural full-rate motion or production suitability.

Delivery, checksums, full decoding and actual LangSmith media attachment retrieval were verified separately from visual or listening quality. Comparable native experiments were checked for the completed pairs and 100-image arms. LangSmith's preview/comparison UI was not inspected.

## Public-copy verification and attribution

Original media remain retained privately. PNG copies were losslessly re-encoded with ancillary metadata removed, with decoded pixel bytes compared against each source. Video and FLAC containers were remuxed without metadata and without re-encoding; encoded stream hashes were compared against each original. Source and public-copy SHA-256 hashes are both recorded. Every public output was fully decoded; `publication-audit.json` retains the original 123-file checks. `append-audit-qwen-turbo-ten.json` records the ten additional PNG checks and proves original media/records were preserved. Video contact sheets are derived JPEGs, not replacement original clips.

Prompts, measurements and observations are supplied as benchmark evidence. Model and companion terms remain separate from Packet Foundry's Apache-2.0 code license. In particular, FLUX.1 dev uses its non-commercial model license; YuE2 has CC-BY-NC-4.0 model terms; Qwen Image 2.1 uses its research terms; FLUX.3 components link their upstream FLUX Kommunity and companion terms. The model-source and license links in the dataset identify the relevant upstream material. This archive does not grant a new blanket license to model weights or override upstream conditions on their outputs.


## Qwen Turbo follow-up and choosing seed variants

Ten further local images use the distinct Unsloth Qwen Image 2.1 Turbo Q4_K_M checkpoint, the official eight-step ManualSigmas schedule, CFG1 and Euler. The first two share a brief with seeds 42/43; eight subsequent requests use new briefs. All ten outputs are retained, including visible material, geometry or constraint errors. Independent original-image visual review is separate from successful relay, decoding and actual LangSmith attachment/experiment checks.

The production cache finding is supported by all 50 prior FLUX pairs: every second request reused both conditioning outputs while sampling again. Excluding initial loading, native medians were 70.05 s for the new brief and 38.49 s for its next seed. The Qwen Turbo pair took 320.246 s then 95.234 s; its first included model loading, so that delta combines startup/model and encoding reuse. `seed-variant-cache-findings.json` separates these scopes, includes the actual measured gaps and maps all 50 FLUX pairs to unchanged public records. No five-variant batch was tested or extrapolated as measured.

Turbo model terms: Qwen research license; not Apache-2.0. https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo/blob/fc13ba1986fcb905854dc8de2f52eee395f49b5b/LICENSE Workflow attribution and MIT notice: `qwen-turbo-workflow-NOTICE.md`. Existing model/media license limitations remain in force; no weights are redistributed.


## Matched Qwen Turbo Q4/Q8 follow-up

Ten actual Q8 outputs reuse the exact original Q4 briefs/seeds/order. The earlier 133 generated files and 140 attempt records remain byte-for-byte in their original per-attempt records/media/workflows. All ten new outputs have independent original-PNG visual notes, real acknowledgement/decoding/attachment checks and native comparable ten-case experiments. qwen-turbo-q4-q8-comparison.json separates the initial request, next seed of the same brief and eight new-brief population; it maps each side-by-side pair to its public records and exact timing/cache context. Q4 originals were not regenerated. No automatic aesthetic superiority, isolated quantization speedup, whole-VRAM residency or five-variant timing is claimed. `append-audit-qwen-turbo-q8.json` records the additional PNG checks and preserved baseline hashes. Existing workflow MIT notice and model/media terms remain in force; no weights are redistributed.