A ComfyUI process breakdown
A study on giving skin back its photographic truth without giving up creative freedom, built on my multi-camera workflow with Flux 2 Klein and Qwen and tested on a head sheet generated in Ideogram 4.
Drag the line to compare the original sheet with the final version.
Ideogram 4 delivers the most convincing skin I have ever worked with. That quality, however, comes with a limit. When I raised the resolution, the character stayed identical: the seed changed, the eye color and hair tone changed, but the architecture of the face remained intact. It was always the same person.
The release of Flux 3 changed that. In under seven hours I had the weights installed and the first generations ready, and the difference was obvious: it lets you deform, stylize and design features that don't give away the model behind them. The only thing missing was what Ideogram did so well: the skin.
Until then, reaching 4K meant going through SeedVR2. It raises quality consistently up to about 2K, but starting from a 1024 px base it can't hold the same level, and in no case does it create skin texture. With Flux 3 generating natively at 4K, realism stops being a question of resolution and becomes a finishing step: a filter.
The challenge was to build that filter on top of the workflow I already used for detailing and multi-camera work, with Flux 2 Klein and Qwen. The sheet used in the tests below was generated in Ideogram 4.
Two open-weight models with different strengths. A summary of what each maker publishes and of what I saw in practice.
Released June 2026
Released October 15, 2026
What I saw in Ideogram 4. Each workflow mode (turbo, default and quality) tends toward its own type of face, and quality, the highest-resolution mode, converges on a recurring face. My reading is that this reflects how the open weights were trained at different resolutions. Ideogram doesn't detail that process, so I treat it as a working hypothesis, not a technical fact.
Where things stand now. Ideogram 4.5 is still available only as a hosted model, and the release of its weights is anticipated. Meanwhile, Flux 3 has just arrived, and the community's first tests already show body sheets working.
flux-2-klein-9b, with the Flux 2 VAEqwen_3_8b_fp8mixed through the TextEncodeQwenImageEditPlus node, a pairing I have used with Klein for a whileThe first version of the workflow took in three views and returned four: an extra head appeared, cropped at the edge, and the panels lost their proportions. Realism, the whole point, barely showed up.
Reading the graph and the metadata revealed three overlapping causes. The sheet went in whole, at about seven megapixels, far beyond the range where Klein works comfortably, so the model started replicating content to fill the space. The prompt asked to preserve the original style, the exact opposite of the goal. And two nodes inherited from the Qwen-Image-Edit template were interfering with the generation; disabling them was the first fix.
The solution was to treat each view as a separate image. I started with a small Python script that finds the real dividers between panels and crops each view losslessly, sorting the results into folders automatically for every new sheet.
Then I brought the same principle into ComfyUI. Image Tile splits the sheet, each view goes through Klein, and Image Untile rebuilds the set at the new resolution. The script stays on hand for sheets with irregular panels.
For both splitting and reassembly, overlap is set to zero so that no view bleeds into its neighbor.
This was the central discovery. A prompt written in advertising language, with rim light, cinematic contrast and an editorial look, brought hard shadows and oily skin. Asking to keep the beard made the beard grow. Asking for smooth skin turned the character into a doll.
In a model without a negative prompt, there is no neutral instruction: whatever you name tends to get drawn. Even the word “marks” turned into blotches.
The final version doesn't describe skin, light, age or gender, so it works for any character, from a child to an elderly person. It lists what must stay and asks only for more definition in what is already there.
Upscale this photograph of the same person. Keep everything exactly as it is: face, identity, age, expression, skin, hair, eyes, clothing, pose, camera angle, framing, lighting, exposure, colors and background. The only change is higher resolution and sharper definition of the details already present: individual hair strands, eyebrows and eyelashes, the iris pattern, and the fabric weave. A faithful, sharp, high-resolution copy of the original photograph.The negative field stays empty.
With the same prompt and the same seed, I compared the result with and without LoRA. Without it, Klein filled the missing detail with inventions: acne on the cheek, a mole on the cheekbone. With the realism and detail LoRAs, it refines what is already there and respects every mark in the original.
Moving Realism Slider from v1 to v2 made the profile even more faithful, with less stippling and less redness.
Denoise determines how much the model redraws. At 0.4, the image comes back almost untouched, grain and all. At 0.7, irises, eyelashes and hair gain definition without the face changing. At 1.0, detail peaks, at the cost of a slight retouch: a line on the forehead, a shadow under the eyes.
I chose 1.0. This sheet will be seen on its own, without the original beside it, and richness of detail outweighed absolute fidelity. When the sheet needs to be an exact copy, 0.7 is the sweet spot.
