Best Image Editing Models for Upscaling in 2026: Flux2-Klein-9B-Consistency and FLUX.2-klein-4B

Quick Answer

Flux2-Klein-9B-Consistency (dx8152) is the top pick as of September 2026 because its release is the newest among the eligible candidates, with no public benchmark yet [1][3][4][5][6][7][8][9]. In order, the ranking is Flux2-Klein-9B-Consistency (dx8152), FLUX.2-klein-4B (black-forest-labs), FLUX.2-klein-base-4B (black-forest-labs), AnyPose (lilylilith), Qwen-Image-Edit-2511-Lightning (lightx2v), Qwen-Image-Edit-2511 (Qwen), BFS-Best-Face-Swap (Alissonerdx), and Qwen-Edit-2509-Multiple-angles (dx8152) [1][3].

Key Takeaways

• dx8152’s Flux2-Klein-9B-Consistency ranks first because its March 5, 2026 release is the newest among the admitted candidates; the position does not establish superior upscaling quality.[8]
• Ranking note: no public benchmark scores any candidate, so the rule gives: newest release first, then downloads; every entry has “no public benchmark yet,” and downloads serve only as a tiebreak.[1][3][4][5][6][7][8][9]
• Scope: this ranking admits only image-to-image models from labs with a published paper or leaderboard record, or other publishers above 10,000 downloads in the last 30 days, using Hugging Face’s image-to-image pipeline tag.[1][2][3][4][5][6][7][8][9]
• black-forest-labs’ FLUX.2-klein-4B and FLUX.2-klein-base-4B share a January 14, 2026 release date; their respective 403,340 and 279,742 downloads over the last 30 days, measured September 25, 2026, determine their order.[3][4] Both use Apache-2.0 licensing.[3][4]
• Plan hardware separately from downloads: Qwen’s Qwen-Image-Edit-2511 lists 57.7 GB of BF16 weights,[1] while lightx2v’s Qwen-Image-Edit-2511-Lightning lists 107.7 GB of full-precision weights.[7] Neither file-size figure alone establishes GPU memory or system RAM requirements.
• Check the license of the specific release: Alissonerdx’s BFS-Best-Face-Swap uses MIT,[6] while lilylilith’s AnyPose and dx8152’s Qwen-Edit-2509-Multiple-angles use Apache-2.0.[9][5] Licensing does not establish upscaling performance.

How do local image editing models compare on specifications and public upscaling benchmarks?

Which local image editing models should you consider for upscaling?

1. Flux2-Klein-9B-Consistency

Flux2-Klein-9B-Consistency by dx8152 ranks first because its publication date, 2026-03-05, places it ahead of the other eligible releases under the recency rule [8]. The ordering basis is newest release first, then downloads. Its benchmark status is “no public benchmark yet,” so the position does not establish superior upscaling quality.

The repository lists 0.7 GB of full-precision weights and an Apache-2.0 license [8]. That weight total does not establish the hardware needed for a complete local workflow. A verified GPU memory requirement, system RAM requirement and tested hardware configuration are unavailable for this ranking; confirm those requirements before allocating hardware.

Use the model as an exploratory upscaling candidate, checking whether outputs preserve the input’s details. The practical caveat is unverified task performance: the release date supports its position, but no public upscaling score supports choosing it for image fidelity.

2. FLUX.2-klein-4B

FLUX.2-klein-4B by black-forest-labs ranks second under the ordering rule: newest release first, then downloads.[3][4][8] Its publication date, 2026-01-14, matches black-forest-labs’ FLUX.2-klein-base-4B.[3][4] The tie is resolved by 403,340 versus 279,742 downloads over the last 30 days, measured as of 2026-09-25.[3][4] Download counts determine placement here, not demonstrated upscaling quality.

The model has 3.9 billion parameters, lists 23.7 GB of BF16 weights, and uses the Apache-2.0 license.[3] For local deployment, the weight listing does not establish a GPU or system-memory requirement. Verify the complete inference configuration before choosing hardware; do not treat the listed file size as a VRAM requirement.

For upscaling, consider it a candidate for controlled evaluation on your own images. Check whether enlarged outputs preserve text, edges, and fine textures before adopting it. The caveat is straightforward: no public benchmark yet establishes its upscaling performance, so its ranking is a shortlist position rather than a measured quality advantage.

3. FLUX.2-klein-base-4B

FLUX.2-klein-base-4B by black-forest-labs ranks third under the ordering rule “newest release first, then downloads.” [3][4] Its Hugging Face publication date matches FLUX.2-klein-4B at 2026-01-14, but its 279,742 downloads trail that model’s 403,340 over the last 30 days, as of 2026-09-25. [3][4] The position reflects release timing and a popularity tiebreak; no public benchmark yet establishes its upscaling quality.

The model has 3.9 billion parameters, lists 23.7 GB of BF16 weights, and uses the Apache-2.0 license. [4] For local hardware planning, distinguish the listed weight size from runtime memory requirements. A specific GPU or system-RAM capacity cannot be recommended from those specifications alone; measure memory use with your intended image dimensions and workflow before choosing hardware.

Use it as a candidate for local upscaling evaluation when Apache-2.0 licensing suits your project. [4] The caveat is unverified upscaling performance: check detail preservation and unwanted changes on representative images before adopting it.

4. AnyPose

AnyPose by lilylilith occupies fourth place under the newest release first, then downloads ordering, based on its Hugging Face publication date of December 25, 2025.[3][4][8][9] AnyPose has no public benchmark yet, so its position does not establish an advantage in upscaling quality.

The repository lists full-precision weights totaling 0.6 GB and an Apache-2.0 license.[9] AnyPose recorded 23,114 downloads over the preceding 30 days as of September 25, 2026.[9] Download activity provides context for adoption, but does not demonstrate detail recovery or image fidelity.

For local use, GPU, VRAM and system-RAM requirements remain unverified; the listed weight size alone cannot establish which hardware will run the complete workflow. Treat AnyPose as an exploratory candidate for your own upscaling evaluation. The practical caveat is the absence of a public upscaling benchmark: assess output against your original images before relying on it for production work.

5. Qwen-Image-Edit-2511-Lightning

Qwen-Image-Edit-2511-Lightning by lightx2v ranks fifth under the ordering rule of newest release first, then downloads.[7][9][1] Published on December 22, 2025, it follows lilylilith’s AnyPose and precedes Qwen’s Qwen-Image-Edit-2511 by release date.[7][9][1] Its position reflects release timing; there is no public benchmark yet establishing its upscaling quality against the other candidates.

The repository lists full-precision weights totaling 107.7 GB and an Apache-2.0 license.[7] Treat that weight total as a storage planning input, not a measured VRAM requirement. A specific GPU configuration or system RAM requirement cannot be established from the published specifications, so confirm the complete pipeline’s memory needs before choosing local hardware.

Consider it for local upscaling evaluation when permissive licensing matters.[7] Compare representative images against your originals before adopting it. The practical caveat is unverified upscaling performance: neither the Lightning name nor its ranking establishes faster execution or better detail preservation.

6. Qwen-Image-Edit-2511

Qwen-Image-Edit-2511 by Qwen ranks here because the ordering is newest release first, then downloads, placing its December 17, 2025 release behind the newer candidates [1][3][4][7][8][9]. For upscaling, its status is no public benchmark yet; the position does not establish a quality advantage over another model.

The model lists 20.4 billion parameters, 57.7 GB of BF16 weights and an Apache-2.0 license [1]. Local hardware requirements remain unspecified: the listed weight size does not establish runtime GPU memory or system RAM needs. A GPU recommendation would require a defined runtime, precision and image resolution, followed by memory measurements.

Consider it for a local upscaling evaluation when Apache-2.0 licensing suits your project [1]. Check whether enlarged outputs preserve text, edges and fine textures before adopting it. The caveat is the missing public upscaling benchmark: treat detail preservation and reconstruction quality as properties to validate, rather than capabilities established by its ranking.

7. BFS-Best-Face-Swap

BFS-Best-Face-Swap by Alissonerdx ranks here under the ordering rule “newest release first, then downloads.” Its publication date is 2025-11-07.[6] The model has no public benchmark yet, so its position does not establish upscaling quality. Its 131,229 downloads in the last 30 days, as of 2026-09-25, indicate adoption rather than measured performance.[6]

The repository lists 12.1 GB of full-precision weights and an MIT license.[6] Local hardware requirements remain unspecified: the weight-file size does not establish a GPU memory requirement or total system RAM requirement. A hardware purchase cannot be justified from that figure alone; confirm runtime requirements before committing to a deployment.

For engineers considering face swapping within an upscaling workflow, treat the model as an evaluation candidate. Check whether outputs preserve identity and fine detail on representative images. The caveat is straightforward: its ranking provides no measured evidence of resolution recovery or faithful detail reconstruction.

8. Qwen-Edit-2509-Multiple-angles

Qwen-Edit-2509-Multiple-angles by dx8152 ranks eighth because its publication date, 2025-10-31, places it last in this candidate set under the ordering rule: newest release first, then downloads.[5] Its upscaling status is no public benchmark yet; the position reflects release order, not a measured difference in detail recovery.

The repository lists full-precision weights of 0.2 GB and an Apache-2.0 license.[5] Hugging Face recorded 179,178 downloads over the preceding 30 days as of 2026-09-25.[5] Download activity indicates adoption, but does not establish upscaling quality or determine its position ahead of newer releases.

For local deployment, GPU memory and system RAM requirements remain unverified. The listed weight size should not be treated as a complete runtime memory budget. A practical role is evaluation within an image-editing workflow: compare outputs against the original for preserved texture, edges and identity before adopting it for enlargement. The central caveat is the lack of a public upscaling benchmark to substantiate quality claims.

What hardware requirements can you establish from published model weight sizes?

Published model weight sizes establish storage requirements for the listed weights, but they do not establish total GPU memory, system RAM or a specific GPU recommendation. Hardware planning needs the complete inference configuration; a weight-file total alone cannot establish whether an upscaling workflow fits on your machine.

Qwen-Image-Edit-2511 (Qwen), published by Qwen, lists BF16 weights of 57.7 GB.[1] Black Forest Labs’ FLUX.2-klein-4B (black-forest-labs) and FLUX.2-klein-base-4B (black-forest-labs) each list BF16 weights of 23.7 GB.[3][4] Treat those figures as storage budgets for the published weights, rather than measured VRAM requirements.

The full-precision weight totals also vary across publishers. lightx2v’s Qwen-Image-Edit-2511-Lightning (lightx2v) lists 107.7 GB,[7] while Alissonerdx’s BFS-Best-Face-Swap (Alissonerdx) lists 12.1 GB.[6] Neither figure establishes inference speed, output resolution or the memory needed during an upscaling operation.

The remaining entries list smaller full-precision weight packages: dx8152’s Flux2-Klein-9B-Consistency (dx8152) lists 0.7 GB,[8] lilylilith’s AnyPose (lilylilith) lists 0.6 GB,[9] and dx8152’s Qwen-Edit-2509-Multiple-angles (dx8152) lists 0.2 GB.[5] Package size alone does not establish whether additional model weights are required. Before choosing hardware, confirm the complete model dependencies, runtime precision and memory measurements at your intended output resolution. A GPU purchase cannot be justified from these file sizes alone.

Which licenses apply to local image editing models?

The listed local image editing models use Apache-2.0 or MIT, depending on the release [1][3][4][5][6][7][8][9]. For a local deployment, match the license to the exact repository you plan to download.

Qwen’s Qwen-Image-Edit-2511 carries Apache-2.0 [1]. The separately published Qwen-Image-Edit-2511-Lightning from lightx2v also carries Apache-2.0 [7]. Both entries therefore have the same stated license, despite their different publishers [1][7].

The FLUX.2-klein-4B and FLUX.2-klein-base-4B models from black-forest-labs both carry Apache-2.0 [3][4]. Choosing between those releases does not change the stated license [3][4]. Keep the complete repository name in your deployment notes so the base release and the other release remain distinguishable.

The Flux2-Klein-9B-Consistency and Qwen-Edit-2509-Multiple-angles models from dx8152 carry Apache-2.0 [8][5]. AnyPose from lilylilith also carries Apache-2.0 [9]. BFS-Best-Face-Swap from Alissonerdx carries MIT [6].

For an engineering handoff, record the model repository, downloaded revision and applicable license together. Review the license text before redistributing weights or incorporating a model into a shipped product. Treat the license choice separately from your upscaling evaluation: use the license record to guide deployment review, and evaluate output quality against your own images and acceptance criteria.

How should you choose an upscaling model with no public benchmark yet?

Choose an upscaling model with no public benchmark yet by testing your own images, checking its license, and measuring resource use on your hardware.

Use the same input images and target dimensions for each trial. Include faces, text, fine textures, and compressed images relevant to your work. Compare whether the output preserves identity, lettering, edges, and texture without adding unwanted detail. Record processing time and peak memory alongside your visual assessment.

Treat download counts as a tiebreak, not proof of upscaling quality. The ranking basis is “newest release first, then downloads”; dx8152’s Flux2-Klein-9B-Consistency takes the first position because its publication date is later than every other candidate’s, not because of demonstrated upscaling performance.[8][1][3][4][5][6][7][9] The ranking admits only image-to-image models from labs with a published paper or leaderboard record, or other publishers above 10,000 downloads in the last 30 days.[5][6][7][8][9]

Check the complete workflow before choosing hardware. Qwen’s Qwen-Image-Edit-2511 lists 57.7 GB of BF16 weights under Apache-2.0.[1] black-forest-labs’ FLUX.2-klein-4B lists 23.7 GB of BF16 weights under Apache-2.0.[3] Use those figures as weight-storage information; request measured memory requirements for your intended configuration before buying a GPU.

Keep the performance verdict explicit: “no public benchmark yet.” Choose the model that preserves the details you need while meeting your measured runtime, memory, and licensing requirements.

Frequently Asked Questions

Which model leads the upscaling shortlist?

Flux2-Klein-9B-Consistency (dx8152), published by dx8152, ranks first because its release date of 2026-03-05 makes it the newest eligible candidate.[8] Its status is “no public benchmark yet,” so that position does not establish superior upscaling quality. Use the shortlist to choose candidates for local evaluation; a release-date ranking cannot tell you which model will preserve detail in your images.

How is the ranking decided?

Ranking note: no public benchmark scores any candidate, so the order is newest release first, then downloads.[1][3][4][5][6][7][8][9] Scope: this ranking admits only image to image models from labs with a published paper or leaderboard record, or other publishers above 10,000 downloads in the last 30 days (Hugging Face pipeline tags: image-to-image).[1][2][3][4][5][6][7][8][9] Downloads break release-date ties; they do not demonstrate upscaling quality.

What is the complete model order?

The order is Flux2-Klein-9B-Consistency (dx8152);[8] FLUX.2-klein-4B (black-forest-labs);[3] FLUX.2-klein-base-4B (black-forest-labs);[4] AnyPose (lilylilith);[9] Qwen-Image-Edit-2511-Lightning (lightx2v);[7] Qwen-Image-Edit-2511 (Qwen);[1] BFS-Best-Face-Swap (Alissonerdx);[6] and Qwen-Edit-2509-Multiple-angles (dx8152).[5] Every entry has the status “no public benchmark yet.” Read this sequence as a release-based evaluation order, not a measured comparison of upscaling results.

How much GPU memory do I need to run these models locally?

A complete GPU-memory requirement is not established. FLUX.2-klein-4B has 3.9B parameters:[3] parameter storage at 4-bit precision would be 1.95 GB, estimated (params x 0.5 bytes).[3] Qwen-Image-Edit-2511 has 20.4B parameters:[1] the equivalent figure is 10.2 GB, estimated (params x 0.5 bytes).[1] Both calculations cover parameter storage only; neither establishes complete runtime VRAM, system-RAM requirements, or compatibility with a particular GPU.

Which licenses apply to these models?

BFS-Best-Face-Swap (Alissonerdx) lists the MIT license.[6] Every other candidate in this shortlist lists Apache-2.0.[1][3][4][5][7][8][9] For a local deployment, record the license attached to the exact repository you download and review its terms for your intended use. Keep that check separate from image-quality evaluation: a license label does not establish how well a model will upscale your images.

Which model has a public benchmark proving its upscaling quality?

Every candidate has the status “no public benchmark yet”; no shared public score establishes an upscaling winner.[1][3][4][5][6][7][8][9] Qwen’s Qwen-Image Technical Report is dated 2025-08-04,[2] but that publication date is not a dated upscaling result for this shortlist. For a practical comparison, use the same input images and target dimensions, then inspect text, edges, textures, and unintended changes before choosing a model.

Sources

1. Qwen/Qwen-Image-Edit-2511 model card (Hugging Face) — 2026-09-25
2. Qwen-Image Technical Report — 2025-08-04
3. black-forest-labs/FLUX.2-klein-4B model card (Hugging Face) — 2026-09-25
4. black-forest-labs/FLUX.2-klein-base-4B model card (Hugging Face) — 2026-09-25
5. dx8152/Qwen-Edit-2509-Multiple-angles model card (Hugging Face) — 2026-09-25
6. Alissonerdx/BFS-Best-Face-Swap model card (Hugging Face) — 2026-09-25
7. lightx2v/Qwen-Image-Edit-2511-Lightning model card (Hugging Face) — 2026-09-25
8. dx8152/Flux2-Klein-9B-Consistency model card (Hugging Face) — 2026-09-25
9. lilylilith/AnyPose model card (Hugging Face) — 2026-09-25