<h1 id="qwenimagevae_liquid1127">qwenimageVAE_liquid1127</h1><p><code>qwenimageVAE_liquid1127</code> is a lightly tuned Qwen Image VAE based on <code>qwenimageVAE_liquid1087</code>. It was adjusted with Krea2-style image generation in mind, but it can also be used with other Qwen Image compatible models, such as Anima and other Qwen-based workflows.</p><p>This VAE was tuned with a focus on:</p><ul><li><p>stronger and cleaner black line rendering</p></li><li><p>slightly warmer pale-orange tonal gradation</p></li><li><p>preserving color richness without making the image look oversaturated</p></li><li><p>reducing unwanted pink / magenta fringing around high-contrast edges</p></li><li><p>maintaining brightness, contrast, and smooth reconstruction quality</p></li><li><p>improving stability for detailed anime-style illustrations and semi-realistic outputs</p></li></ul><p>The goal of this VAE is not to drastically change the model’s style, but to subtly improve how decoded images handle linework, color boundaries, warm highlights, and edge artifacts.</p><h2 id="recommended-use">Recommended use</h2><p>This VAE is intended for Krea2 / Qwen Image based workflows, especially when generating:</p><ul><li><p>detailed anime-style illustrations</p></li><li><p>character portraits with fine linework</p></li><li><p>images with orange, red, pink, purple, and blue color transitions</p></li><li><p>high-contrast lineart</p></li><li><p>images where magenta edge halos or pink fringing are a problem</p></li><li><p>outputs where the original VAE feels slightly color-thin or weak in black line definition</p></li></ul><h2 id="tuning-notes">Tuning notes</h2><p>This VAE was tuned using reconstruction-based losses, not a full style-training approach.</p><p>The purpose of the tuning was to make small decoder-side numerical adjustments while preserving the original behavior of <code>qwenimageVAE_liquid1087</code> as much as possible.<br />The encoder was frozen, and only a very small decoder-tail portion was updated.</p><p>In this training setup, the trainable parameters were limited to decoder output-tail modules such as:</p><pre><code>decoder.conv_out
decoder.norm_out
decoder.head
decoder.conv2</code></pre><p>This means the VAE was not broadly retrained.<br />Instead, only the final output-side part of the decoder was adjusted so that line clarity, luminance balance, local contrast, color separation, and magenta-fringe behavior could be slightly corrected without changing the VAE too aggressively.</p><h2 id="training-setup">Training setup</h2><p>The main training settings were:</p><pre><code>RESOLUTION = 640
BATCH_SIZE = 1
MAX_STEPS = 2000
SAVE_EVERY = 200

LR = 1e-8
WEIGHT_DECAY = 0.0
GRAD_CLIP = 0.03

TRAIN_DECODER_ONLY = True
TRAIN_TAIL_ONLY = True
MIXED_PRECISION = "bf16"
LATENT_CLAMP = 3.0</code></pre><p>The VAE weights themselves were kept in <code>float32</code> during training for stability, while <code>bf16</code> autocast was used for the forward pass.<br />The encoder was executed under <code>no_grad</code>, and decoder-only optimization was used to reduce VRAM usage and avoid destabilizing the latent distribution.</p><h2 id="loss-weights">Loss weights</h2><p>The final loss was a weighted combination of several reconstruction and image-quality terms:</p><pre><code>W_RECON_L1 = 1.00
W_MULTI_SCALE = 0.14
W_EDGE = 0.10
W_LUMA = 0.14
W_CONTRAST = 0.10
W_COLOR_RICHNESS = 0.02
W_ANTI_PINK = 0.16
W_KL = 1e-6</code></pre><p>The actual total loss was calculated as:</p><pre><code>total_loss = (
    W_RECON_L1 * loss_recon
    + W_MULTI_SCALE * loss_ms
    + W_EDGE * loss_edge
    + W_LUMA * loss_luma
    + W_CONTRAST * loss_contrast
    + W_COLOR_RICHNESS * loss_color_richness
    + W_ANTI_PINK * loss_anti_pink
    + W_KL * loss_kl
)</code></pre><p>Since this was decoder-only tuning with a frozen encoder, the KL term was effectively kept as a placeholder rather than being used as a major training objective.</p><h2 id="reconstruction-loss">Reconstruction loss</h2><p>A standard L1 reconstruction loss was used as the main anchor:</p><pre><code>loss_recon = F.l1_loss(recon_f, target_f)</code></pre><p>This keeps the decoded image close to the original training image and prevents the VAE from drifting too far from the base VAE behavior.</p><h2 id="multi-scale-reconstruction-loss">Multi-scale reconstruction loss</h2><p>A multi-scale L1 loss was used to preserve both small details and larger tonal structures.</p><pre><code>def multiscale_l1(recon, target):
    loss = F.l1_loss(recon, target)
    r = recon
    t = target
    for _ in range(3):
        r = F.avg_pool2d(r, kernel_size=2, stride=2)
        t = F.avg_pool2d(t, kernel_size=2, stride=2)
        loss = loss + F.l1_loss(r, t)
    return loss / 4.0</code></pre><p>This helps the VAE match not only pixel-level details, but also broader gradients, soft shading, and overall tonal structure.</p><h2 id="edge-preservation-loss">Edge preservation loss</h2><p>A Sobel-based edge loss was used to improve line stability and high-frequency detail reconstruction.</p><pre><code>edge_r = sobel_mag(y_r)
edge_t = sobel_mag(y_t)
loss_edge = F.l1_loss(edge_r, edge_t * EDGE_TARGET_GAIN)</code></pre><p>The edge target was slightly strengthened:</p><pre><code>EDGE_TARGET_GAIN = 1.03</code></pre><p>This was intended to make black lines, eyelashes, hair strands, clothing folds, and other fine illustration details slightly cleaner without making the result overly sharp.</p><h2 id="luminance-preservation-loss">Luminance preservation loss</h2><p>A luminance loss was used to reduce unwanted brightness shifts.</p><pre><code>def luminance(x):
    return 0.299 * x[:, 0:1] + 0.587 * x[:, 1:2] + 0.114 * x[:, 2:3]

loss_luma = F.l1_loss(y_r, y_t)</code></pre><p>This helps keep the tuned VAE from making images too dark, too bright, or overly contrast-heavy compared with the source reconstruction.</p><h2 id="local-contrast-loss">Local contrast loss</h2><p>A local contrast/statistical contrast term was used to prevent the image from becoming too flat.</p><pre><code>def contrast_stat(y):
    return y.flatten(2).std(dim=2)

loss_contrast = F.l1_loss(contrast_stat(y_r), contrast_stat(y_t))</code></pre><p>This helps preserve separation between shadows, highlights, and lineart.</p><h2 id="color-richness-loss">Color richness loss</h2><p>A small color-richness loss was used to preserve useful chroma separation.</p><pre><code>def chroma_richness_stat(x):
    mean = x.mean(dim=1, keepdim=True)
    chroma = (x - mean).pow(2).mean(dim=1, keepdim=True).sqrt()
    return chroma.mean(dim=(2,3))</code></pre><p>The target chroma richness was slightly increased:</p><pre><code>COLOR_RICHNESS_TARGET_GAIN = 1.03
rich_r = chroma_richness_stat(recon_f)
rich_t = chroma_richness_stat(target_f) * COLOR_RICHNESS_TARGET_GAIN
loss_color_richness = torch.relu(rich_t - rich_r).mean()</code></pre><p>This was not meant to simply oversaturate the image.<br />The goal was to prevent the decoded result from becoming color-thin, especially in red, orange, purple, blue, and pale warm gradients.</p><h2 id="anti-pink-anti-magenta-fringe-loss">Anti-pink / anti-magenta fringe loss</h2><p>A custom anti-fringe loss was used to suppress unwanted magenta or pink halos around high-contrast edges.</p><pre><code>def anti_pink_fringe_loss(recon, target):
    y_t = luminance(target)
    e_t = sobel_mag(y_t)
    thr = e_t.mean(dim=(2,3), keepdim=True) * 1.20
    edge_mask = (e_t &gt; thr).float()

    r = recon[:, 0:1]
    g = recon[:, 1:2]
    b = recon[:, 2:3]

    magenta_excess = torch.relu((0.5 * (r + b)) - g - 0.03)
    return (magenta_excess * edge_mask).mean()</code></pre><p>This loss only focuses on edge regions detected from the target image.<br />It penalizes cases where red and blue become too strong compared with green near edges, which often appears visually as pink or magenta fringing.</p><p>The intention was not to remove normal pink, purple, or red colors from the image.<br />Instead, it was designed to reduce unwanted magenta-tinted edge artifacts while preserving intentional warm and colorful areas.</p><h2 id="stability-measures">Stability measures</h2><p>Several safety measures were used to avoid NaN collapse and unstable VAE behavior:</p><pre><code>MAX_BAD_STEPS = 200
LATENT_CLAMP = 3.0
GRAD_CLIP = 0.03
optimizer = torch.optim.AdamW(trainable, lr=LR, weight_decay=WEIGHT_DECAY, eps=1e-4)</code></pre><p>Latents were clamped before decoding:</p><pre><code>latents = torch.nan_to_num(
    latents,
    nan=0.0,
    posinf=LATENT_CLAMP,
    neginf=-LATENT_CLAMP
).clamp(-LATENT_CLAMP, LATENT_CLAMP)</code></pre><p>Decoded images were also sanitized before loss calculation:</p><pre><code>recon_f = torch.nan_to_num(
    recon_f,
    nan=0.0,
    posinf=1.0,
    neginf=-1.0
).clamp(-2.0, 2.0)</code></pre><p>If non-finite gradients were detected, that step was skipped instead of updating the weights.</p><h2 id="key-format-and-saving">Key format and saving</h2><p>The base VAE used ComfyUI/WebUI-style Qwen Image VAE key names.<br />For training, the weights were loaded into <code>diffusers.AutoencoderKLQwenImage</code> using a shape-and-order based key mapping.</p><p>After training, the model was saved back in the original ComfyUI/WebUI Qwen Image VAE safetensors key format, so it can be used directly in ComfyUI-style workflows.</p><p>The saved weights were exported as <code>bfloat16</code> safetensors.</p><h2 id="notes">Notes</h2><p>Because this is a subtle VAE tuning, the difference may be easiest to see in side-by-side comparisons using the same seed, same model, same prompt, and same latent image, changing only the VAE.</p><p>Look especially at:</p><ul><li><p>black line clarity</p></li><li><p>hair strand separation</p></li><li><p>eye highlights</p></li><li><p>edge halos</p></li><li><p>red / orange / purple color boundaries</p></li><li><p>smoothness of skin and fabric gradients</p></li><li><p>whether fine details become muddy or remain readable</p></li></ul><h2 id="disclaimer">Disclaimer</h2><p>This VAE is an experimental fine-tuned VAE for visual comparison and workflow testing.<br />Results may vary depending on the base model, sampler, prompt, LoRA, and image style.</p>

Files with SHA256: 54650bd81d070ba02449ffe33e5957ba752fda223f3a7ea3ae3038243e097677