diff --git a/runtime/ui/menu_art.cpp b/runtime/ui/menu_art.cpp
index 3a623ca0..ad7f31f8 100644
--- a/runtime/ui/menu_art.cpp
+++ b/runtime/ui/menu_art.cpp
@@ -134,6 +134,11 @@ static int upload_rgba(Art &a, const uint8_t *rgba, uint32_t w, uint32_t h, uint
     }
     memcpy(map, chain, (size_t)bytes);
     free(chain);
+    // The batch flushes its own staging buffers. This one is caller-owned and can be noncoherent.
+    if (vmaFlushAllocation(c->vma, sa, 0, bytes) != VK_SUCCESS) {
+        vmaDestroyBuffer(c->vma, sb, sa);
+        return art_free(a), 0;
+    }
     VkCommandBuffer cb = vk_upload_begin();
     VkImageMemoryBarrier2 b = {VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER_2};
     b.srcStageMask = VK_PIPELINE_STAGE_2_NONE;
diff --git a/runtime/vk_astc.c b/runtime/vk_astc.c
index daa4e5b9..04bda92b 100644
--- a/runtime/vk_astc.c
+++ b/runtime/vk_astc.c
@@ -483,13 +483,19 @@ static VkBuffer scratch_get(VkCtx *c, VkCommandBuffer cb, VkDeviceSize need, int
         VmaAllocationCreateInfo aci = {.usage = VMA_MEMORY_USAGE_AUTO_PREFER_DEVICE};
         VkBuffer b;
         VmaAllocation a;
-        if (vmaCreateBuffer(c->vma, &bi, &aci, &b, &a, NULL) != VK_SUCCESS) {
+        VkResult r = vmaCreateBuffer(c->vma, &bi, &aci, &b, &a, NULL);
+        if (r == VK_ERROR_OUT_OF_DEVICE_MEMORY && bi.size > need) {
+            // The usual scratch leaves room for later textures. This upload can still fit without the spare bytes.
+            bi.size = need;
+            r = vmaCreateBuffer(c->vma, &bi, &aci, &b, &a, NULL);
+        }
+        if (r != VK_SUCCESS) {
             plat_mutex_unlock(&g_lock);
             return VK_NULL_HANDLE;
         }
         // the old one may still be read by its command buffer: freed once the batches recorded so far completed
         if (S.buf && S.dev == c->device) vk_defer_buffer(S.buf, S.alloc);
-        S.buf = b, S.alloc = a, S.size = size, S.dev = c->device;
+        S.buf = b, S.alloc = a, S.size = bi.size, S.dev = c->device;
         n_scratch_new++;
         vk_texstat_time(VK_TS_SCRATCH, t0);
     } else {
diff --git a/runtime/vk_memory.c b/runtime/vk_memory.c
index 1c17fd59..5574d9c6 100644
--- a/runtime/vk_memory.c
+++ b/runtime/vk_memory.c
@@ -3679,6 +3679,12 @@ void *vk_staging(VkDeviceSize size, VkBuffer *buf, VmaAllocation *alloc) {
     VkBuffer b;
     VmaAllocation a;
     VkResult r = vmaCreateBuffer(c->vma, &bi, &aci, &b, &a, &info);
+    if (r == VK_ERROR_OUT_OF_DEVICE_MEMORY && aci.pool) {
+        // The staging pool can grow only in 64 MiB blocks of its one memory type. A smaller allocation may
+        // still fit in a default block or another host-visible type. Keep the same mapping requirements.
+        aci.pool = NULL;
+        r = vmaCreateBuffer(c->vma, &bi, &aci, &b, &a, &info);
+    }
     if (r != VK_SUCCESS) rt_fatal("vk: staging buffer of %llu bytes failed (%d)", (unsigned long long)size, (int)r);
     vk_texstat_add(VK_TS_STG_NEW, 1, bi.size);
     *buf = b;