Use Case

This tool probably serves a very niche use case, but it was one I was looking for specifically. Used to generate an initial image with FLUX1D and then upscale with tile stitching to better overlap and reduce artifact/distortions from a singular 4x-Upscaler node.

I wanted to generate cycle-able desktop wallpapers and struggled to find good variety at the 32:9 image format that weren't just regular ultra wide resolutions upscaled with artifacts and zoom impact.

Using the tiled KSampler in Step 3 massively reduces artifacts created by the 4x-Upscaler step, especially on detailed images or images with specific subjects/people. Tiling this also helps reduce OOM for lower RAM/VRAM machines.

Using Other Models: As a general note, you CAN wire other models into this (as a base model) instead of Flux if you so choose, but out of testing Flux handles the most image styles while also generally being the best at respecting the 32:9 image format. Testing with Anima/ZIT/Illustrious all had successes, but the format is distinct enough to make them less frequent, so I would recommend splitting the workflow in this case, doing Step 2/3 separately in one pass and starting with a Load Image not (making it a solo I2I workflow as opposed to T2I). This will matter far less with more abstract concepts and patterns, but for subjects the difference is night and day.

Update: I've also wired up an Anima (Diff Model) version of this for upload as well. You can swap ZIT into this (with appropriate updates to CLIP/VAE). The Anima version does pretty poorly with photo realism, even with appropriate checkpoints due to the output resolution, but does work well for artistic/anime style images.

Overview

• Step 0 (Pre):

  • Set your base settings, models, image seed, and latent image size (2560x720 is used here to generate final output of 5120x1440 for 32:9 desktop wallpapers)

• Step 1 (First Image):

  • Set your Positive and Negative prompts using Flux best practices and choose any LoRAs you want to use. Initial generation image (2560x720) will be saved as "Base_#" in output folder.

• Step 2 (Upscale):

  • Image is upscaled, pick your upscale model/method according to your needs but use a 4x (either AnimeSharp or UltraSharp).

• Step 3 (Tiled Stitch and Upscale KSampler)

  • Second lighter model is loaded specifically to assist the Ultimate SD Upscale node. Set a best practice prompt based on the model family. Select the model here based on your use case (match photorealistic to photorealistic models, animated/anime models to animated/anime images, etc.), tiled upscale is run with overlap and seam fixing at minimal denoise (as low image impact as possible), final cleaned/upscaled image is output to the same folder as "Post_#" in output folder.

Baseline Models Used - FLUX1D (Swappable)

• Step 1

  • flux1-dev-Q8_0.gguf (GGUF)

  • t5xxl_fp8_e4m3fn.safetensors (CLIP)

  • clip_l.safetensors (CLIP)

  • ae.safetensors (VAE)

• Step 2

  • 4x-AnimeSharp.safetensors (Upscaler) OR

  • 4x-UltraSharp.safetensors (Upscaler)

• Step 3

  • Choose a light AIO model (VAE / Clip Included) that suits your use case best

Baseline Models Used - Anima (Swappable)

• Step 1

  • anima_baseV10.safetensors (Diff Model)

  • qwen_3_06b_base.safetensors (CLIP)

  • qwen_image_vae.safetensors (VAE)

• Step 2

  • 4x-AnimeSharp.safetensors (Upscaler) OR

  • 4x-UltraSharp.safetensors (Upscaler)

• Step 3

  • Choose a light AIO model (VAE / Clip Included) that suits your use case best

Prompts

For Step 1 use best practices for Flux prompts, understand that you're not creating a 1:1 or 3:2 image, so adjust your prompt to best fit a horizontal 32:9 style for best results. Direction, alignment notes, etc., should all match widescreen image generation best practices.

For Step 3 your positive and negative prompts should be far less generative and directive, more reinforcement on stylistic choices, and prompts should follow best practices for the model family selected (Illustrious, SDXL, Z-Image, and so on) when applied.

Details/Notes

FLUX1D is used here specifically for better output at the awkward starting resolution (2560x720) than most other generators, and gives more flexibility in output types than a lot of SDXL/Illustrious line models. Z-Image or ZIT theoretically works here as well, but I found via testing Flux has the overall best output for moving on to later steps.

Output is clean at 5120x1440 (I have a Samsung Odyssey G9) and should work well for 1440 dual screen setups where your desktop wallpaper spans.

You have control over the original GGUF load (I'm using flux-1-dev-Q8_0.gguf for my quant, you can come down on this depending on your hardware), the Upscale Model used (pairing 4x-UltraSharp or 4x-AnimeSharp for realism or animated), and the final upscale model used for the stitched tile pass.

For the stitched tile pass specifically the prompt is non-generative, but does have impact. Model selection also matters here (match your model to your first image type), I found in testing that Illustrious line "realism" models create a lot more artifacting than purpose built photo realism models, if you're doing a photograph style image I would recommend swapping this to a realistic SDXL or ZIT AIO for Step 3. I tested a BF16 ZIT AIO for a few passes on photographic style wallpaper images and it worked phenomenally, but it will substantially increase overall workflow time and RAM pressure, so for general use a lighter weight/faster model would be my recommendation.

For cartoon/animated/anime style wallpapers my total workflow time averaged ~100-140 seconds (initial image generation in Flux, upscale, and then tiled stitch/blend). Using the BF16 ZIT AIO bumped this up to ~9.5 minutes (but the quality was notable).

My Specs:

• CachyOS Linux

• ComfyUI + ComfyUI_Manager

• NVidia MSI RTX 5080 Gaming Trio OC (16GB)

• AMD Ryzen 7 9800X3D

• 64GB G.Skill DDR5 CL28 RAM

Version Complete Workflow Time w/ Hardware:

• Diff Model Version (Anima Base) - ~120-150 Seconds

• GGUF Version (FLUX1D) - ~100-140 Seconds

When to use which version?

• Lower Hardware Specs

  • Anima - Far less RAM/VRAM pressure overall, less cut/quantized base models fit entirely within 8/12/16GB of RAM. Flux will require heavy quants even on moderately powerful consumer hardware/general workstation PCs that are well built

• Photo Realism

  • Flux - By a mile. Better support for the wider image format (stretching beyond even the normal 3:2 format in this case to 32:9. The best Anima realism models are still pretty bad

• Anime/Animated Characters

  • Anima - Doesn't get much easier than saying "Frieren from Frieren" to handle 95% of a character generation, efficient token use and ease as a result

• Lora Access (General)

  • Tossup - Flux is fairly well supported with general Loras, and Anima will be more specifically trained with Lora's for specified output sizes (which this workflow is actively fighting against in general), but there are a lot of Anima LoRAs and more being added daily

• Lora Access (Anime/Character Specific)

  • Anima - Pretty easy to train, community focused on this specific thing, growing daily. They exist for Flux, so don't completely ignore it, but this is really easy to source material for for Anima

• Lora Access (Photo Realism)

  • Flux - Again by a mile. The Anima photo realism LoRAs are ... coming along, but at time of publish they're still pretty questionable at best. Amplify this for every aspect of weirdness outputting at this resolution format creates. Go Flux if this is your goal.

• Newer to generation

  • Anima - Less custom setup required, easier levers to tweak, currently very popular so easy to find support/help within CivitAI and other AI communities. This becomes a Tossup if you understand how ComfyUI_Manager works and can setup GGUF requirements w/o issue

This workflow does require multiple custom nodes, I recommend using with ComfyUI_Manager