The full report is 514MB so not everything is included but let me know if you want to see more and I'll try to find away to host the full report.
Phase 1 + Phase 2 - 14 entries - all crops at identical image-relative coordinates - final resolution 2496x3648
How to read this report
Scores are visual judgements made at 200-400% magnification on nearest-neighbour crops, not computed metrics. SSIM/PSNR appear near the bottom as supporting evidence only: they reward doing nothing and would rank a blurry resample above any genuine reconstruction, which is precisely the failure this benchmark is designed to avoid.
The original source row is the 1248x1824 input expanded 2x with nearest-neighbour. It is not an upscale and is not scored. It exists so you can see exactly what detail was present before any method ran, which is the only way to tell recovered detail from invented detail.
Category results
Best fidelity - Z-Image 4x upscaler
Best added detail - Forge Neo reference
Best clean detail (detail without overcooking) - A. Z-Image -> Krea 2 RAW tiled
Best skin - A. Z-Image -> Krea 2 RAW tiled
Best hair - Forge Neo reference
Best eyes - A. Z-Image -> Krea 2 RAW tiled
Best lace / fabric - Forge Neo reference
Best jewelry - B. SeedVR2 -> Krea 2 RAW tiled
Lowest overcooking - Z-Image 4x upscaler
Best performance - SeedVR2 3B INT8
Best overall balance - A. Z-Image -> Krea 2 RAW tiled
Best cleanup of a noisy first pass - E. SeedVR2 -> Z-Image
Why
Best fidelity goes to Z-Image on the strength of structural agreement (SSIM 0.9287 vs source) and unchanged tone. Branch A scores the same 9 for fidelity and is the better image; Z-Image wins the narrow category because it changes least.
Best added detail goes to Forge on raw reconstruction alone - it resolves more genuinely new lash, iris, lace and hair information than anything else here. It wins this category and loses the artifact ones.
Best hair and best lace/fabric also go to Forge for the same reason: nothing else resolves individual hairs or thread-level weave as far. Branch A is the choice if the grain that comes with it is unacceptable.
Best eyes goes to branch A rather than Forge. Forge resolves more lash and iris detail, but it grains the eyelid heavily; A gets most of the structural gain with clean skin around it, which is the better eye overall.
Best jewelry goes to branch B: SeedVR2's metal crispness with a large part of its noise removed.
Best performance is SeedVR2 at 26.3 s, an order of magnitude faster than anything in Phase 2 and still a credible standalone result if its skin speckle is tolerable.
Best overall balance is branch A, which leads or ties on eight of the ten positive axes while keeping both risk scores low.
Best cleanup of a noisy first pass is the new branch E: it is the only method that removes a large fraction of SeedVR2's speckle and chroma noise while keeping its iris, lace and chain structure, and it does so in 93 s.
Phase 3 — the missing hybrid orderings
Phase 3 added SeedVR2 -> Z-Image (branch E) and Z-Image -> Flux -> SeedVR2 (branch F). The third requested pipeline, SOURCE -> SeedVR2 -> Flux tiled, was already completed in Phase 2 as branch D and was reused rather than re-run. Stage-by-stage high-frequency figures are in Missing_Hybrids/Comparison/hf_stage_analysis.json; the focused eight-way sheet is Missing_Hybrids/Comparison/focused_contact_sheet.png.
1. Does SeedVR2 -> Z-Image retain SeedVR2's detail advantage while improving its artifact behaviour?
Largely yes. HF falls 12.4% with only 30.5% of blocks gaining, cheek speckle is largely removed and the chroma blotching in blacks is almost eliminated, while iris fibre, lace cells and chain definition survive. That is outcome (A) - preserves useful detail, cuts artificial HF - not (B) or (C). The caveat is that eyelid speckle survives and the finest strands soften slightly, so it improves SeedVR2 markedly without reaching plain Z-Image's cleanliness.
2. Does Flux -> SeedVR2 produce more useful detail than either model alone?
No. It produces more high-frequency energy than either (27.27 against 16.26 for Flux and 25.48 for SeedVR2), but much of it is pseudo-detail: beaded hair strands, etched lace rims and injected hard edges. Real gains are limited to metal edge definition.
3. Does Flux -> SeedVR2 become overcooked at 200-400%?
Yes, unambiguously. It is the most overcooked entry in the benchmark - HF amplified across 70% of blocks, up to 4.3x in hair, with visibly dotted strands and chroma speckle returning to the blacks.
4. Is SeedVR2 -> Flux cleaner than Flux -> SeedVR2?
Yes, decisively. Branch D reduces SeedVR2's HF by 12.9% with only 9.3% of blocks gaining, while branch F amplifies by 53% across 70% of blocks. Whichever model runs last sets the artifact profile, and SeedVR2 is a poor finisher because it amplifies rather than reorganises.
5. Does either missing hybrid challenge Z-Image -> Krea RAW d=0.14 as best overall balance?
No. Branch A stays top on composite. Branch E lands a solid third overall and beats A on nothing; branch F ranks near the bottom. A remains the quality choice.
6. Is any new hybrid close enough to Krea RAW to replace the ~1300 s pipeline for normal production?
Not as a like-for-like replacement of branch A - E scores lower on skin, eyes and detail gain. But E is a genuinely useful new option at 93.5 s, 14x faster than A, and it is the best answer when the first pass is noisy. For clean inputs plain Z-Image at 61 s remains the better fast default; E earns its place specifically as a SeedVR2 repair path.
Per-method observations
Forge Neo reference
Resolves the most genuinely new information of any entry: individual eyelashes, radial iris fibre, thread-level lace weave and many separate hair strands where the source had undifferentiated mass.
Pays for it with a global fine grain. At 400% the cheek and eyelid carry a diagonal cross-hatch weave and dirt-like speckle that is not in the source, and dark strands break into granular dots.
Worst near-black behaviour under a 2.6-gamma shadow lift: dense purple/green chroma speckle across the satin. Highest artifact load of any method.
Highest broadband high-frequency energy relative to its structure (spectral peak/median 14.3 vs 44-48 for the others), which is the signature of added noise rather than added structure.
Impressive fit-to-screen, visibly noisy at 200-400%. This is the classic case the brief asks to penalise.
Z-Image 4x upscaler
The most structurally faithful result (SSIM 0.9287 vs source, by far the highest) with tone essentially unchanged.
Cleanest deep blacks in the benchmark: under shadow lift the satin stays smooth with clear material separation while Forge and SeedVR2 both break into chroma noise.
Conservative rather than inert - it does resolve lace cells, chain links and lash separation - but leaves the iris plain and renders hair as smooth broad strands rather than individual hairs.
Carries a faint diagonal cross-hatch lattice on the eyelid and iris. This originates here, and is inherited by every branch built on it.
Fine vellus hairs on the cheek render as broken dotted segments rather than continuous strokes.
SeedVR2 3B INT8
Best iris reconstruction of any first pass: irregular, organic crypt structure rather than smooth grey.
Metalwork and chain links are the crispest in Phase 1, though pushed toward a polished CG look.
Disqualifying skin defect: dense bright specks scattered across the eyelid and cheek that read as glitter makeup and exist nowhere in the source. Clearly visible at 400% and still detectable at 200%.
Coarse orange and green chroma blotching in near-blacks under shadow lift, worse in character (if less dense) than Forge's fine grain.
Highest high-frequency energy of any entry (25.48), driven by speckle rather than structure.
Fastest method in the whole benchmark at 26.3 s.
Krea2 Turbo refine d=0.08
Structurally this is not a 2x reconstruction at all. It refines at the native 1248x1824 and reaches 2496x3648 by lanczos resample, so no method in this group can resolve detail that was not already present.
High-frequency energy of 5.60 against 14.9-25.5 for every genuine upscaler quantifies the softness - it is the blurriest entry in the benchmark by a wide margin.
Introduces warm brown blotches and specks into near-monochrome skin, and lifts shadows (13.99% of pixels below level 8 versus 15.08% in the source), flattening the black-on-black separation the image depends on.
Scores well on fidelity and artifact control largely because it changes so little.
Krea2 Turbo refine d=0.14
Same lanczos-to-target limitation as d=0.08; more denoise buys drift, not resolution (SSIM vs source falls to 0.7284).
Warm colour contamination is stronger than at d=0.08 and shadow lift is greater still.
Krea2 Turbo refine d=0.20
Most deviation from source in Phase 1 (SSIM 0.6768) with no corresponding detail gain - the worst trade in the benchmark.
Strongest warm/sepia contamination and the most lifted blacks (12.97% below level 8 versus 15.08% in the source).
Krea2 Turbo d=0.14 -> Z-Image (hybrid)
The only Phase 1 Krea Turbo branch that performs a real 2x reconstruction, because Z-Image does the upscaling after the refine (HF energy 16.89, in line with Z-Image's 16.90).
Inherits the warm colour contamination from the Turbo refine, so it is measurably less tonally faithful than plain Z-Image while costing 112 s against 61 s.
Eyelid takes on a waxy, slightly plastic quality and lashes show mild smearing.
Superseded by Phase 2 branch A, which is the same idea done in the correct order with the genuine RAW model.
A. Z-Image -> Krea 2 RAW tiled
The strongest example in the benchmark of added detail without overcooking. It consolidates Z-Image's broken, dotted vellus hairs into continuous strokes, sharpens lash separation and gives the iris real fibre structure that Z-Image leaves flat.
Crucially its high-frequency energy DROPS relative to its parent (14.43 vs 16.90) while visible structure increases. It is reorganising existing information, not injecting noise - the opposite of the Forge/SeedVR2 signature.
Tonally near-identical to its parent (18.13% of pixels below level 8 vs 18.03%), so the black-on-black separation Z-Image protects is preserved.
Flat skin at 800% is as clean as Z-Image: no pores, no speckle, no plastic sheen.
Only notable defect is a faint, softly directional micro-texture in the iris, visible at 400%. A spectral test does not support calling it a regular lattice (peak/median 44.7, statistically indistinguishable from Z-Image's 44.0 and the source's 47.6), so it is low-contrast rather than structured.
Adds fine satin weave and stitch detail in dark fabric, with slight chroma speckle appearing only under aggressive shadow lift.
Verified seam-free at production size: boundary z-excess -0.96 to +1.17 against a non-boundary baseline of 5.4 (vertical) and 15.8 (horizontal).
By far the most expensive branch at 1310.7 s.
B. SeedVR2 -> Krea 2 RAW tiled
Meaningfully repairs SeedVR2: high-frequency energy falls from 25.48 to 18.03 and the eyelid glitter speckle is substantially reduced.
Keeps SeedVR2's strengths - crisp metalwork, well-defined lace cells, strong iris - making it the best jewelry result in the benchmark.
Does not fully remove the inherited speckle; scattered specks remain visible on skin at 400-800%.
Inherits SeedVR2's contrast push and its coarse chroma blotching in near-blacks, so material separation stays mid-pack.
1275.0 s.
C. Z-Image -> Flux 2 Klein tiled
Behaves as a contrast-and-smoothing pass rather than a detail pass. Lashes come out softer and the iris flatter than branch A from the identical input.
Crushes blacks: 21.63% of pixels fall below level 8 against 18.03% in its Z-Image parent, and p99 rises to 247. Smooth satin gradients flatten out, which is costly in an image built on black-on-black material distinction.
Amplifies the eyelid cross-hatch lattice inherited from Z-Image - the most visible instance of that artifact anywhere in the benchmark.
Loses faint secondary hair strands that are present in the parent.
No seams and no geometric drift; the alignment concern raised during inspection did not survive a phase-correlation check (dy=0, dx=0).
Cheap at 303.7 s - roughly a quarter of branch A.
D. SeedVR2 -> Flux 2 Klein tiled
The most effective cleanup of SeedVR2 in the benchmark. Flux's smoothing tendency is a liability on the clean Z-Image input but an asset here: it removes most of the glitter speckle and much of the coarse chroma blotching (HF energy 25.48 -> 22.20).
Still crushes blacks (20.83% below level 8 vs 17.23% for its parent) and flattens gradients.
Retains more of SeedVR2's useful crispness in lace and metalwork than branch C retains of Z-Image's character.
Fastest Phase 2 branch at 213.2 s, six times quicker than branch A.
E. SeedVR2 -> Z-Image
Answers the question it was built for: Z-Image does clean up SeedVR2 without gutting it. High-frequency energy falls 25.48 -> 22.32 (-12.4%) and only 30.5% of image blocks gain HF, so this is a net reduction, not a second sharpening pass.
On flat cheek skin at 800% SeedVR2's bright speckle is largely gone - the single biggest artifact fix in the benchmark. The iris keeps SeedVR2's radial fibre structure and the chain and lace cells stay well defined, so the cleanup is not simply blur.
Incomplete on the eyelid: the larger, brighter specks there survive the 2x downsample and re-upscale and remain clearly visible at 400%. Skin is much better than SeedVR2's but not clean like plain Z-Image.
Best black-material result of any SeedVR2-derived branch by a wide margin. Under a 2.6-gamma lift the orange/green/magenta chroma blotching that disfigures SeedVR2's satin is almost entirely removed.
Costs are a slightly deeper black point (19.99% of pixels below level 8 vs 17.23% for SeedVR2) and the highest chroma saturation in the benchmark (9.00), a mild warm push.
Lashes and the finest strands are a little softer than SeedVR2's, which is the price of the 1248x1824 handoff.
Cheap: 93.5 s total (30.3 s SeedVR2 + 63.2 s Z-Image), about 1/14 of branch A.
F. Z-Image -> Flux 2 -> SeedVR2
Overcooked. Ends at HF energy 27.27, the highest of any entry in the benchmark - above SeedVR2 alone (25.48) and Forge (23.51) - having amplified its Flux parent by +53% across 70% of image blocks, reaching 4.3x in the hair.
The clearest pseudo-detail failure found anywhere here: thin bright hair strands are rendered as dotted chains of discrete bright dots where Z-Image, Flux alone and branch E all render the same strands as continuous lines. SeedVR2 also injects hard thin edges that are absent from its parent.
Reintroduces chroma speckle into blacks that its Flux parent had clean, and compounds the black crush - 21.75% of pixels below level 8, the highest contrast (std 90.22) and the lowest median (26) in the benchmark.
Genuinely good in two places: flat cheek skin stays clean (no speckle returns), and metal edges on the chain gain real definition.
Lace becomes crisper but with bright etched rims around the cells rather than resolved weave.
Running SeedVR2 last is the problem, not SeedVR2 itself - it has no clean signal left to reconstruct from and instead amplifies whatever the previous stage produced.
390.8 s cumulative (61.3 Z-Image + 303.7 Flux + 25.9 SeedVR2).
Recommended pipelines
1. Z-Image -> Krea 2 RAW tiled, d=0.14 (branch A) - quality tier
When the output matters more than the clock. ~22 min per image at 2496x3648, ~11.8 GB peak VRAM.
Best overall balance in the benchmark and the only method that adds substantial detail while measurably reducing high-frequency noise.
Tonally faithful to its parent, so it preserves the black-on-black material separation this image depends on.
Cost is the objection, not quality: at 1310.7 s it is 21x SeedVR2 and 4x branch C.
If iris micro-texture ever becomes objectionable, drop denoise toward 0.10 - the preflight showed d=0.10 at SSIM 0.897 vs 0.865 for d=0.14.
2. Z-Image alone - fast, safe tier
Batch work, or any job where predictability beats maximum detail. 61 s per image.
Cleanest artifact profile in the benchmark and the best deep-black handling, at 1/21 the cost of branch A.
Genuinely reconstructs lace cells, chain links and lash separation - it is conservative, not inert.
Accept that the iris stays flat, hair renders as broad strands rather than individual hairs, and a faint eyelid lattice is present.
This is the sensible default, with branch A reserved for hero images.
3. SeedVR2 -> Flux 2 Klein tiled, d=0.15 (branch D) - noisy-input repair tier
When a fast first pass has left speckle or chroma noise that needs cleaning. 26 s + 213 s.
The best available repair path: Flux's smoothing removes most of SeedVR2's glitter speckle and chroma blotching while keeping its crisp lace and metalwork.
Total cost of ~240 s is still under a fifth of branch A.
Accept ~3.6 points more black crush and flatter gradients; do not use this pairing on an already-clean input, where branch C shows the same smoothing becomes pure loss.
Keep branch B instead if jewelry crispness is the priority and the extra 1000 s is affordable.
4. SeedVR2 -> Z-Image (branch E) - noisy-input repair, fast tier
When SeedVR2's speed is wanted but its speckle is not. 93.5 s total, ~10.4 GB peak VRAM.
The cheapest large artifact reduction available: removes most cheek speckle and nearly all chroma blotching in blacks for 93.5 s, 14x faster than branch A.
Keeps SeedVR2's iris fibre, lace cell and chain definition - it is a cleanup, not a blur.
Accept residual eyelid speckle, slightly softer fine strands, about 2.8 points more black crush than SeedVR2 and a mild warm push.
Preferable to branch D as a SeedVR2 repair when black-level fidelity matters, since Flux crushes blacks harder.