Does GPU Change Texture Quality? H100 vs RTX 4090 for Physical Texture LoRA — 312-Run Benchmark

Disclosure: This article was written with access to a Floyo compute environment provided through a partnership arrangement. Discount codes are available on my profile.

Quick background: Japanese painting in high school, art university, then a master's in Museum Studies at a UK university. The current work is making LoRAs that try to capture the physical texture of traditional Japanese craft materials — gold leaf, raden (mother-of-pearl inlay), lacquerwork. 312+ training runs on RunPod RTX 4090 so far.

The question that keeps coming back: can AI learn to reproduce the actual feel of these materials, not just their color?

When access to an H100 came through Floyo, the first test wasn't speed. Everyone already knows H100 is faster. The real question: does the GPU change what the texture looks like? Does hakuashi edge translucency come out differently? Does thin-film interference in raden resolve differently?

That's what this article is about. Not a standard speed benchmark — a texture comparison.

The 312 runs mentioned in the title are my cumulative training background on RTX 4090 — the baseline experience that lets me evaluate what changes (and what doesn't) when moving to H100. The direct comparison uses identical parameters across a focused set of texture LoRAs.

Setup

ItemSpecification Training toolKohya_ss / Flux LoRA Trainer Base modelDreamShaperXL Alpha2 DatasetSHIFUKU Series (gold leaf + raden) Resolution1024×1024 LoRA Rank32 LoRA Alpha16.0 Learning Rate0.0001 (1e-4) OptimizerAdamW 8bit LR Schedulercosine Mixed Precisionfp16 Batch Size1 (RTX 4090) / up to 4 (H100) Epochs10 (5,200 total steps) Platform ARunPod (RTX 4090 24GB, $0.69/hr) Platform BFloyo (H100 NVL, 94GB VRAM)

Same dataset, same epoch count, same seed for output comparison across both platforms.

Platform Specs Comparison

SpecRTX 4090 24GBH100 NVL 94GBRatio VRAM24GB GDDR6X94GB HBM33.9× Memory bandwidth~1,008 GB/s3,938 GB/s3.9× FP16 tensor~330 TFLOPS1,671 TFLOPS5.1× Cost$0.69/hr~$2.50-7/hr (plan dependent)—

3.9× the memory bandwidth. For diffusion-based training, memory bandwidth is the primary bottleneck, so this is a big deal for training speed. More on that later.

A Note on Methodology

An upfront disclosure: RunPod used A1111 (WebUI) and Floyo used ComfyUI. Different UI platforms means the seed implementation differs, so outputs won't be pixel-identical even with the same seed number. This comparison isn't a pixel-diff analysis. It's an evaluation of texture expression based on 312 runs of experience.

As covered in the environment comparison article, the visual differences below come from the UI platform (A1111 vs ComfyUI), not the GPU. Quality transfers fine between platforms.

Shared Parameters

ParameterValue CheckpointDreamShaperXL Alpha2 SamplerDPM++ 2M SchedulerKarras CFG Scale7 Size1024 × 1024

Per-Pattern Parameters

• Crane (NoLoRA):
  No LoRA / Steps 40 / Seed 1267967603

• Cat (NoLoRA):
  No LoRA / Steps 35 / Seed 2176997588

• Gold Leaf Crane:
  GOLDLEAF_v1 strength 0.60 / Steps 20 / Seed 3897294094

• Raden Cat:
  RADEN_PEARL_v2 strength 0.80 / Steps 35 / Seed 2176997579

A Word on Speed

Keeping this short because it's not what this article is about.

Yes, H100 is faster. The 94GB VRAM and 3,938 GB/s bandwidth mean noticeably quicker training, especially when pushing the batch size up. With RTX 4090's 24GB, going above batch 2 is risky (OOM territory). With H100, batch 4 is comfortable. Bigger batches smooth the loss curve and reduce overfitting — which matters a lot for texture LoRAs, because overfitting in this case means learning noise patterns instead of material properties.

But the question that actually matters here: does training on faster hardware change what the output looks like? That's next.

Texture Comparison

This is the part that matters, and the part most benchmark articles skip entirely. Speed is just a number. Evaluating texture requires knowing what you're looking at.

The whole research is built on one observation: AI is great at adding things — colors, details, complexity. It's bad at reproducing the absence of smoothness that makes real materials feel real. Gold leaf isn't gold-colored paint. It's beaten, pressed, aged, cracked. The texture comes from physical processes, and teaching a model to capture that means getting it to subtract the digital smoothness, not add more detail.

For deeper reading on texture evaluation:

• The truth behind "slippery" AI textures (Japanese — explores why AI defaults to smooth surfaces)

• The "too many colors" problem (Japanese — why AI adds colors instead of simulating optics)

Baseline (No LoRA)

First, outputs with no LoRA applied. This separates any GPU-related variance from the LoRA's own effect.

Gold Leaf Crane — NoLoRA

RunPod (A1111 + RTX 4090):

A white crane dominates the upper frame, flanked by pine trees in a classic byōbu (folding screen) composition. Strong contrast. The gold tone leans warm amber, and panel joints are visible. Dense, weighty, muscular — reminiscent of the Kanō school in Japanese painting.

Floyo (ComfyUI + H100):

Floyo (ComfyUI + H100) put a darker crane lower in the frame, thick pine trunk on the right, blue mountains in the background. The gold is brighter and cooler. More atmospheric depth and breathing room. More Maruyama-Shijō school — naturalistic, less imposing.

ーーーーーーー

Raden Cat — NoLoRA

RunPod(A1111 + RTX 4090):

RunPod: tabby cat sitting upright, staring directly at the viewer. Gold background with big blue-green lotus leaves, pink flower lower left. The fur rendering is detailed to the point of feeling three-dimensional — more miniature painting than illustration.

Floyo(ComfyUI + H100):

Floyo: same tabby, completely different background. A large lunar halo behind the cat, deep navy sky with scattered golden particles, pink lotus everywhere. Dreamlike. The fur detail is just as good, but the whole image has a softer, more ethereal tone.

What the Baseline Tells Us

No quality gap between the two. Both pull the full capability out of DreamShaperXL.

But the vibe is clearly different. RunPod outputs lean toward high contrast, dense composition, heavy color layering. Floyo outputs lean brighter, more spacious, softer light. The same pattern appeared in the environment comparison article — described as "oil painting vs. watercolor." Same thing here. This is the A1111 vs ComfyUI sampling pipeline doing its thing, not the GPU.

Gold Leaf LoRA

LoRA: SHIFUKU_GOLDLEAF_v1 | Strength: 0.60 | Steps: 20 | Seed: 3897294094

Three things to evaluate in gold leaf output: "is it gold-colored."

Hakuashi (箔足) — the edge where leaf meets ground. Real gold leaf gets pressed onto lacquer or sizing, and at the edges it thins out so the red-brown ground bleeds through. A clean, uniform gold border means the model hasn't learned the material. It's just painting things gold.

Diffuse reflection variance — gold leaf is covered in microscopic wrinkles from the beating process. Light scatters unevenly across the surface. Uniform reflection means metallic paint, not leaf. This is the hardest property to get right in AI because the training photos need macro-level shots that capture micro-level texture.

Aging cracks — the nikawa (animal glue) that fixes gold leaf shrinks over time, and old leaf develops a distinctive crackle pattern. Whether the LoRA picks this up depends on the training data and on training stability, which is where batch size (and VRAM) plays an indirect role.

Results

RTX 4090 output:

RTX 4090 produced a landscape composition — crane by water, pine, rocks, mist. Very sansuiga (山水画). Gold is present but the ink-wash gradients hold it back. The overall feeling is subdued, aged — like a screen that's been in a temple for 200 years. The leaf texture is restrained, more "patina" than "gold." A splash of red maple as the only warm accent, and the generated seals look plausible.

H100 output:

H100 went in a completely different direction. Simpler composition — just crane and pine — but the gold ground dominates the image. The wrinkles from the beating process across the whole surface. Light catches unevenly, exactly the way real gold leaf behaves. If the RunPod version is an aged screen, this one feels like freshly beaten leaf — vivid, almost raw.

Same LoRA, same strength. But the A1111 path emphasizes the aging, while the ComfyUI path brings out the physical texture of the leaf itself.

Think of it like pasting the same gold leaf onto coarse linen vs. smooth silk — the underlying surface changes what you notice.

Raden LoRA

LoRA: SHIFUKU_RADEN_PEARL_v2 | Strength: 0.80 | Steps: 35 | Seed: 2176997579

Raden is the hardest one. The material is thin-sliced abalone shell set into lacquer. The iridescence comes from thin-film interference — same physics as a soap bubble. The color shifts with viewing angle. It's not that multiple colors exist on the surface at once.

Most AI models get this wrong because they add rainbow colors instead of simulating the optics. The real test: can the model produce a surface where color is implied by how light behaves rather than directly painted?

Results

RTX 4090 output:

RTX 4090 output looks like a top-down view of an actual raden makie lacquer tray. Circular medallion composition. Six or seven abalone shell fragments arranged around the cat, each one throwing its own iridescence — pink, blue, green holographic light, contained within each shell piece. The cat's fur picks up a faint rainbow tint too. Overall it reads as decorative craftwork — raden treated as an object you can point at and say "that's the raden."

H100 output:

H100 output is totally different. The cat is reclining, surrounded by lotus and gold. But the raden effect isn't contained in individual shell pieces anymore — it's running through the cat's body. Soft pink-to-green interference across the white fur. The cat looks like it's wearing the light. Less "craft object" and more "optical phenomenon."

Same LoRA at the same strength, resolving in two different directions. The A1111 path concretizes the effect as discrete objects. The ComfyUI path diffuses it across the whole scene.

Likely a broader tendency of the two sampling pipelines, not specific to raden, but more test cases would be needed to confirm it.

The Technical Part (Why Quality Doesn't Change)

If you've been paying attention, you might be thinking: "the images are different though." They are. But the difference is in style, not in quality. And the style difference tracks to A1111 vs ComfyUI, not RTX 4090 vs H100.

Here's why:

Image generation (inference) is a math problem. Start from noise (determined by the seed), apply a series of denoising calculations (determined by the sampler, steps, and CFG), arrive at an image. Same inputs, same math, same result. The calculator brand doesn't change the answer to an addition problem.

Can GPUs introduce any numerical difference? Technically, yes — floating-point rounding can vary slightly between GPU architectures. RTX 4090 is optimized for FP16, H100 for BF16. The order of additions in convolution operations can also differ. But in practice, inter-GPU floating-point error measured by SSIM (Structural Similarity Index) is typically above 0.999. Invisible to the eye.

• Precision types: RTX 4090 excels at FP16; H100 excels at BF16. Both are 16-bit, but the decimal rounding behavior differs subtly.
  

• Computation paths: Within convolution operations, the order of additions can vary. 1+2+3 and 3+1+2 are the same for humans, but in floating-point arithmetic, the accumulation of rounding errors can diverge.

The differences in this article — composition, color palette, atmosphere — are way bigger than rounding error. They come from A1111 and ComfyUI handling noise initialization, scheduler internals, and VAE decoding differently.

Running ComfyUI on RunPod would produce Floyo-like results. Running A1111 on Floyo would produce RunPod-like results. The GPU isn't the variable.

Bottom Line

FactorRTX 4090 24GB (RunPod)H100 NVL 94GB (Floyo)Recommendation Training speedBaselineFaster (3.9× bandwidth)H100 for time efficiency VRAM headroom24GB (tight for batch>2)94GB (batch 4+ comfortable)H100 for training stability Setup time30–60 min (first time)Near zeroFloyo for beginners Cost$0.69/hrPlan-dependentRTX 4090 for cost Output qualityEquivalentEquivalentNo difference

GPU choice doesn't affect texture quality in inference. Not for baseline images, not for gold leaf, not for raden. The exact patterns trained over 312+ runs were tested, and the outputs are equivalent.

The micro-scale texture expressions that texture LoRAs specialize in don't care what GPU rendered them.

The stylistic difference between A1111 and ComfyUI is real, and interesting — RunPod outputs lean heavier, denser, more object-level. Floyo outputs lean brighter, more diffused. But this is a creative choice, not a quality hierarchy. And it can be compensated with LoRA strength adjustments.

• RunPod (A1111): higher contrast, denser composition, texture concretized at the object level
  

• Floyo (ComfyUI): brighter tones, spatial breathing room, texture diffused across the scene

The important takeaway: a LoRA trained on RTX 4090 reproduces the same texture quality when deployed on H100, and vice versa. The problem being solved — getting AI to capture the physical feel of gold leaf and raden — is a dataset design problem and a human evaluation problem. Not a compute problem.

Pick the GPU on speed, cost, and how much time you want to spend on setup. Not on quality.

Where H100 Earns Its Keep

Training, not inference.

H100 NVL's memory bandwidth is roughly 4× RTX 4090. The 94GB VRAM is also about 4×. Diffusion model training hammers VRAM every single step, so the bandwidth advantage translates directly to speed. The value isn't "better images faster." It's "more experiments in the same time window." For someone with the eye to evaluate texture, that speed becomes a force multiplier.

Starting out? RTX 4090 is fine. Doing systematic exploration — testing strength in 0.1 increments, training on 100+ image datasets, comparing batch size effects on loss curves? H100 saves real time.

Want zero-setup? Floyo lets you start training immediately instead of spending the first hour debugging environment issues. Optimizing cost? RunPod's community cloud pricing is hard to beat once the setup is dialed in.

If you prefer zero-setup environments: Floyo is worth trying — you spend your first session training, not troubleshooting.

If you're optimizing cost at scale: RunPod's Community Cloud pricing is hard to beat once you have your environment dialed in.

Related Resources

Try the textures yourself:

• SHIFUKU Series on CivitAI — free texture LoRAs (gold leaf, raden, crushed gems, lacquer, and more) to download and test

Deeper reading on texture evaluation (Japanese):

• The truth behind "slippery" AI textures — Why AI image surfaces all look the same

• The "too many colors" problem in raden output — Why AI adds colors when it should simulate optics

Full comparisons (Japanese):

• RunPod vs Floyo: Full Environment Comparison for Texture LoRA Researchers — Setup, cost, and workflow differences

• Does the GPU Change Texture Quality? Same-Seed Comparison — Detailed Japanese version of this article

TextureLoRA Lab | Shitsukan
Japanese painting → Art studies → Museum Studies MSc (Merit, UK) → AI engineer
Researching the problem of texture absence in AI image generation.
The core insight: AI is good at addition, bad at subtraction. Texture is a subtraction problem.